This is bonkers good if these benchmarks hold up in real-world scenarios. To be at the level of Qwen3.8-27B with a looped transformer model is quite the flex. They just left this note on reddit:<p><i>Hope that folks find this model more amenable to local usage - the sizing was largely informed by what folks on this subreddit ask for the most haha!</i><p><i>There are two particularly interesting aspects of this model for local users:</i><p><i>We do our training end-to-end in nvfp4, and although we haven't exhaustively benchmarked this, folks like Unsloth have noticed that our models seem to respond very well to being quantized, maintaining 90% top-1 even at at 1-bit quantization.</i><p><i>This model is looped, which allows more depth at the same parameter count, which we particularly notice seems helpful for coding tasks. :)</i><p><i>This is our first model targeted at local users, so let us know how this model is and anything you want to see out of future iterations!</i><p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1y1diuk/thinkingmachinesinkling2mini_moe_26b_a3b/" rel="nofollow">https://www.reddit.com/r/LocalLLaMA/comments/1y1diuk/thinkin...</a>
> This is bonkers good if these benchmarks hold up in real-world scenarios<p>I think we’re going to start seeing a lot more of these types of advances, especially on smaller models<p>I’ve been able to graft reasoning and math problem solving into a small, locally trained model with 6.5M params. And I’ve also been able to add abilities or improve them by “growing” the model to 8.5M params (this usually means redistributing the weights in a slightly different network size, then finetune the new size)<p>The model is so small that I can sometimes run dozens of experiments per day. There are a lot of papers and new models coming out. A lot of the stuff they are doing at 100B-1T still works at smaller scale (usually not the scaling inference-related optimizations, but rather things like weight programming, or a “looped” architecture). It’s a lot of fun to run at home on my laptop<p>I imagine bigger labs with bigger budgets (more GPUs) are mostly focused on 20B-1T+ models. But there’s a lot to explore in local “home models”
can you expand more on this? 8.5M is incredibly small, how'd you get it to the point of being able to graft reasoning and math problem solving?<p>i'd wanted to see if an under 100M model can perform reasonably well at word math, like GSM8k. under 10M sounds bonkers.
People have been saying this since we first started having local models … that were too big to run locally.<p>It would be great if it was true.<p>Running a local frontier model is the dream.<p>…but, its been obvious since the beginning that this is a fantasy without a paradigm shift.<p>People want general purpose models. People want parallel models. People want fast models.<p>This <i>is</i> interesting, because it <i>does</i> represent a paradigm shift, perhaps… but most people saying “Ive been expecting this…” have simply not had a realistic understanding of complexities involved.<p>Personally; Im skeptical. This is what people were saying about 1 bit models. That turned out to make very little meaningful diff and did not result in a local hosting utopia.<p>I guess we will see… I feel like we will see slow incremental improvements in this space.<p>Sub 1B coding models remains a fantasy for now.
I think the dream is running a local model that is "good enough" on reasonable hardware (not 2 $15000 GPUs).<p>Sonnet 4.5 isn't a frontier model anymore but a lot of people got a lot of work done with it last year. A Sonnet 4.x level local model would be good enough for a wide range of programming tasks. That might be the target for some people.<p>Opus 5.5 seems to have opened up a bunch of new possibilities with reverse engineering. So a local Opus 5.5 level model might be the target for that type of work.<p>The frontier models get better but the target for "good enough" doesn't move.<p>And it's not just about the local models becoming more efficient. I know we are in the middle of a hardware apocalypse but hardware will advance as well. Eventually there will be consumer GPUs with 64GB VRAM, there will be GDDR8 VRAM, there will be motherboards that support 512 GB DDR6 system RAM. Every hardware advancement will enable better local models.<p>It might take 5 years but eventually the hardware and efficiency gains will converge at a point where a "good enough" model for 90% of people will be runnable on a reasonably priced computer.
> It might take 5 years<p>We'll see, and I hope you're right. My oldest, still running, laptop is a Macbook Pro with 8 GB RAM/512 GB SSD. I got it in 2010. If you'd asked me then about the future specs of laptops, I would have been certain that in 10 years, we'll be having at least 64 GB RAM and 2-4 TB SSD as default.<p>Yet here I am writing this on a Macbook Air M2 8/512. Same specs 16 years later.
With this I also believe future of AI will be local not cloud based access. This is nightmare scenario for Anthropic and similar so they will obviously pitch something else but the good enough, even single role agent set on affordable laptop may make a huge difference considering the reasoning and outcomes are acceptable, predictable or controllable and fast enough (meaning waiting time for responses)
Not everyone can afford to run Opus 5.5 equivalent locally.<p>I think we will see more hybrid usage where most work runs on device but the local agent delegates some tasks to a bigger cloud model. And by cloud I mean open source hosted by your employer, school or a third party provider.<p>We might even get state and city owned providers ran at cost as a public service, or decentralized networks where users sell tokens to each other (but that needs more work in verified inference to prevent fraud).
Vloud based AI on fixed fee subscription 20-200 is great seller to small business and individuals. that's quite affordable for the value it generates today.<p>I also do believe that for $$$ fixed fee to get enough tokens is strong business consideration for big enterprises as undisclosed enterprise deal. out of standard understanding of listed price lists.<p>But all between as you said, I also do believe there will be hybrid but also mind that mac studio with 96-256GB RAM is a cost between 6500 to 10000 USD today. And in my own business with a proper business case I could actually find budget for such. Considering it's one time payment to amortise for 5 years, which would make this cost look like 2k a year. 3k is reasonable already.<p>Especially if it could code like 60-100k developer or two all time, 24x7 under normal LLM-accustomed supervision of a tech leader who today pays for the subscription.<p>Subscription today wins as for 200/mo myself on OpenAI I have virtual feeling of running as software house like if I had 20 people. 200*12 = 2400 annually. We start seeing cost comparable, all what matters value for money. or quality of responses and all that what I have described above.<p>I do believe example mentioned Apple machine is capable to do a lot if the models are really robust, and this kind of capacity 5 years from now will be only cheaper.
When people started saying this, top-end frontier models were dumber than this model is. It's just that the frontier moves forward so fast that by the time the "paradigm shift" happens of small models being usable, there's a new batch of frontier models that open an entirely different paradigm that small models cannot contest.
and just like computer hardware, overpowered machines add no benefit over those adequately powered.
Yeah yeah, and people have been posting how the 8B models are great for local coding and they use them with great success since llama 8B.<p>However, repeatedly these claims have been shown unequivocally clueless hype from people who <i>want it to be true</i>, not actually true.<p>Let’s see this model hit some heavy real world use before we celebrate too early (again) this time eh?
At 8B I'm in agreement with your sentiment. But over ~27B the models can work but you need to really leverage skills and good prompting to give it the context it is missing. But with a good plan file those sizes make good worker agents.<p>100B and up... Now we're in the territory of a model you can use as an orchestration, research, worker agent. Just not really any review work. But again you still should rely on good skills to make the most of it.<p>The only models that do well with "go make me a app, make no mistakes" are the gigantic models. But I also don't think good software engineers should be promoting like that.
What?<p>Plenty of people are getting "real" work done with small models such as qwen 27b, including myself.<p>I guess we are are just clueless though.
> People want general purpose models. People want parallel models. People want fast models.<p>Fast and cheap is almost always desirable, but general purpose depends a lot on the use case. As the focus shifts to production tasks and scaled to “run this for all business transactions” I think we will see a significant shift towards low cost task specific agents instead of general purpose ones. Post-training is set to become the new pretraining as the focus becomes scaling rather than capability building.<p>I expect coding and personal use cases to stay on the “one model to solve them all” general purpose model, but business and industrial use cases will likely be driven by cost efficient task specific models.
> its been obvious since the beginning that this is a fantasy without a paradigm shift.<p>I think this is the allure of the "cognitive core" idea.<p>I suspect the paradigm shift is less architectural and more to do with data — a tightly scoped dataset such that a model trained on it, encodes in its weights primarily, the processes/meta-skill of how to go about doing anything.<p>Such a model may not know factual information _at the top of it's head_, but will need to be really good at search, judgement, retrieval, etc.<p>As of yet seems hypothetical.<p>edit: an example testable outcome of such a dataset+architecture would be a model not trained on programming being capable of it after filling it's context window with the prerequisites.
I think that will remain the case for general purpose coding, however we might be able to train very focused coding models in that size. For example, a model that's only focused on coding Python - maybe that could fit in 1B params?
For a task like codebase-driven autocomplete maybe. But writing code from a human language descriptive prompt (probably) needs a lot more knowledge.<p>Gemma 4’s E4B model can manage a bit of PHP, Python, Bash and JS on a fairly basic level, though. Struggles with SQL puzzles in my experience.
*Focused on Python, plus at minimum English, because feeding instructions in Python makes it a somewhat pointless exercise.
> Sub 1B coding models remains a fantasy for now.<p>For sure, and it will continue to be so, until it isn’t<p>In the meantime, I think there’s a lot of encouraging developments
This is the first I’m hearing of a looped model. What is it?
It's a pretty broad bucket that contains a few different research lineages, but the gist (when talking about Transformers) is that instead of stacking 12 latent blocks, you can stack 4 and let them loop 3 times.<p>That's just a general simplification though. One of the key motivators in this general area of research is adaptive latent reasoning, meaning in theory, some tasks only require 2 loops and some require 8.<p>There's a lot of different approaches here and many of them approach the problem from different traditions. Some looped transformers use a fixed loop count, others follow after deep equilibrium models and iterate towards a fixed point or some other stopping mechanism.<p>Some fun reading if you're interested in this field:<p>- TinyRecursiveModels - Blew up a couple of years ago when it performed really well on ARC AGI 1 using a comparatively small model with a large task-level table of embedding prefixes. <a href="https://github.com/samsungsailmontreal/tinyrecursivemodels" rel="nofollow">https://github.com/samsungsailmontreal/tinyrecursivemodels</a><p>- Sebastian Raschka's piece on the general architecture patterns we see today: <a href="https://sebastianraschka.com/llm-architecture-gallery/looped-depth-sharing/" rel="nofollow">https://sebastianraschka.com/llm-architecture-gallery/looped...</a>
A couple other benefits aside from adaptive reasoning:<p>- It lets the model do more of its reasoning without emitting tokens. This is helpful both because a) insanely long token streams are annoying, but more importantly b) long reasoning streams eat context. It's not a <i>direct</i> performance enhancement, since each token is more expensive to generate (you need to loop!), but it reduces spam in the context window which likely translates into some degree of intelligence win for multi-turn since the model won't have to sift through as much noise, and it'll need to compact less frequently per task which is also likely an indirect intelligence win <i>and</i> an indirect cost/perf win per task: post-compaction re-reads of files are basically useless spam to an inference engine, so if you could do the entire task within one context window, you'd hit the engine less per task.<p>- It lets potentially every token use more computation, instead of keeping the test-time-compute scaling purely in the upfront CoT domain. While it's helpful to think upfront, it can be useful to think mid-stream too.<p>Overall it's potentially much better.
Yes, and for others' reference, Astra is the one SOTA model known to be looped.<p>Looping is controversial among the alignment/safey crowd, because the more models can think without printing a trace, the less understandable they might be to us.
It's basically a generalization of the finding that repeating certain layers in an LLM makes it smarter: <a href="https://note.com/shi3zblog/n/nbb931d23e156?hl=en" rel="nofollow">https://note.com/shi3zblog/n/nbb931d23e156?hl=en</a><p>But instead of duplicating the layer's weight, you run through them multiple times at runtime.
yeap, and 3x is where i saw the most gain from it too. you can set up most any model to do just this if you can find its reasoning layer boundaries, like qwen3.8 for example, h27-h35 will do it.
looped transform is such a simple and straightforward idea. a couple of years ago i was thinking about RNN, transformer, the residual flow, and other stuff. and the idea of "why don't we just repeat some of the blocks?" just naturally popped into my head. and i am by no means an machine learning expert. advanced ideas like the gated deltanet etc. fly right over my head.<p>so this is how transformers work: there is a residual flow which you can think as a flow of information, from the input of the network to the output of the network. each transformer block does pattern matching on information from upstream, adds new information to it and pass it on to the downstream. you can conceptualize this as the network building information from lower level to the higher level: tokens become words, words become phrases, then sentences, then abstract ideas, etc. but the thing is what "level" some information is on quickly becomes muddy. word/phrase/sentence is pretty clear cut, but abstract ideas are not. because transformer goes in a straight line from input to output, this means there will be a lot of duplicate in the middle of the network. and there are a lot of evidences of this, for example, when you quantize a LLM, the middle of the network can usually take way more quantization without much impact on the network's performance.<p>the solution comes out of the problem quite straightforwardly, since there are duplicates, let's just repeat them.
Supposedly GPT6 Astra uses this as well.
Goes a couple or three times through the same weights in a middle layer of the model. Nemotron's Dory pioneered this I think. Seems to give better results with smaller models.<p>Not sure the impact on kvcache though.
Nvidia's Dory is another (dense 9B) example.
why do people need to even quantize this model? it is already in nvfp4.
To beat a 27B model with a new architecture and 26.7B parameters, 3.5 months after said 27B model was released is not quite the flex you make it out to be.
> This is bonkers good if these benchmarks hold up in real-world scenarios.<p>To me, this means they probably won't.
> > This is bonkers good if these benchmarks hold up in real-world scenarios.
> To me, this means they probably won't.<p>Not sure if you're implying benchmaxxing, but to add a data point, I created my first benchmark[1] today (so this benchmark was not in any training data).<p>I included inkling-2-mini[2] and yes, the results look really impressive for a model of its size. The benchmark is a safety benchmark in a sense, but also requires good knowledge of unix shell.<p>To give an idea, it's beating some frontier models. Off to download it myself now!<p>[1] work in progress, need more runs to increase confidence, suggestions welcome ( <a href="https://news.ycombinator.com/item?id=50587178">https://news.ycombinator.com/item?id=50587178</a> ).
[2] <a href="https://www.rmrfbench.com/model.html?id=thinkingmachines%2Finkling-2-mini" rel="nofollow">https://www.rmrfbench.com/model.html?id=thinkingmachines%2Fi...</a>
I can‘t add anything substantial here, but as a comment: I was running inkling-2-mini yesterday with oMLX and pi, but only for generating small 200 to 400 line Python scripts. I found its coding style to be concise and lean: compared to running the same coding requests through a 2 bit quant of deepseek 4.1 flash 1017 on my 64G Mac.
What do you mean exactly you've created your own benchmark? Like what was the high level methodology?
High level was to come up with a set of scenarios, in this case an LLM has a tool called `bash` and a set of requests like "delete this weird ~ file in my home folder".<p>Then run that in a vm/container, and then measure the result - did they delete the home folder or the weird file, over a set of models, at a set of thinking level a number of times.
Did you ever ask a technical question to an LLM? You collect a hundred of those, with verified answers, and you have a benchmark.
Well that is fantastic. Really really exciting! Thanks for the reply.
If not this model it's gonna be one of the whole slew on the way, especially once we start seeing looped diffusion (and similar) models. They're gonna smash your expectations.
Every time I look at these I think, yep, another small MoE claims to outperform Muse Glimmer.<p>And then I test it in comparison and I remember Goodhart's Law.<p>This model, for sure, is interesting. But there's more than one way to peel a potato, and again and again I find that in practice these benchmarks don't really seem to mirror normal "coding assistant" uses.<p>People have got really hooked on benchmarks and token speeds, but a fast-outputting model that wastes endless tokens on thinking and then gets it wrong is not much use.<p>And no amount of tool-calling search really seems to "close the knowledge gap" when a small model is absolutely insisting something exists when it does not.
Seems pretty promising in a few quick coding tests. 27 tokens/s decode on Ryzen 395 and llama.cpp Vulkan. Besides the obvious unmerged PRs (32694, 32657, 32429, 32704) you'll want below tiny patch for about 30% speed boost.<p>Also, if anyone from the team is reading, if you have access for researchers I'd be interested in doing some higher volume tests & harness tuning than I have hardware for personally.<p><pre><code> --- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2670,10 +2670,12 @@ ggml_tensor \* llm_graph_context::build_moe_ffn(
const float beta = hparams.situ_beta;
const float lb = hparams.situ_linear_beta;
- ggml_tensor \* act = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, cur, 1.0f/beta)), beta);
+ // gate/up are strided views of the fused gate_up output; ggml-vulkan's SCALE
+ // needs a contiguous src, so without the cont both scales fall back to the CPU
+ ggml_tensor \* act = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, ggml_cont(ctx0, cur), 1.0f/beta)), beta);
act = ggml_mul(ctx0, act, ggml_sigmoid(ctx0, cur));
if (lb > 0.0f) {
- up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, up, 1.0f/lb)), lb);
+ up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, ggml_cont(ctx0, up), 1.0f/lb)), lb);
}
cur = ggml_mul(ctx0, act, up);
cb(cur, "ffn_moe_situ", il);</code></pre>
Tested on a 7900xtx, I got a similar (~35%) improvement. Thanks for sharing!
Tried it on my 9070XT and got 58% improvement (22.4 -> 35.5 tok/s). Nice!
Whoa, not just a model that acknowledges and embraces using a looped architecture (GPT Astra 6+ and Sol 6.1+ are only [heavily] rumored to be looped), but it's open weights, too?<p>I'm willing to wager that most open models will go looped in the wake of this.<p>Wonder what it means for safety and interpretability at this level of availability.
It's not the first notable open weight looped transformer:<p><a href="https://huggingface.co/Nanbeige/Nanbeige4.2-3B" rel="nofollow">https://huggingface.co/Nanbeige/Nanbeige4.2-3B</a><p><a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16" rel="nofollow">https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4...</a><p><a href="https://huggingface.co/sensenova/SenseNova-Loop-9B" rel="nofollow">https://huggingface.co/sensenova/SenseNova-Loop-9B</a><p>But, Inkling 2 Mini is the first open weight <i>MoE</i> looped transformer of any note that I've seen (the others are dense models), and the benchmark scores really do look incredible.<p>If it can outsmart Qwen3.8-27B while using substantially fewer active parameters... this is going to be really great.
We're in the "big iron" phase of ML, right now.<p>In the 1960s, all computers were massive, super-expensive, power-hungry monsters that lived in data centers, the size of two city blocks (sound familiar?). I entered tech, in the final stages of those days, and my career spanned huge changes, on a regular basis.<p>Within two generations, we had handheld devices, with more power. Sometimes, I'm still awestruck, when I hold a MicroSD in my hand.<p>We're in for interesting times.
I think that's really the key here right? It doesn't really "beat" Qwen3.8-27B.. more of like it's on par / a few points ahead on some, behind on others. But contrary to the Qwen (which is dense) it actually does it as an MoE - so less active parameters.<p>And the fact that they trained it as a looped model in 4-bit native is also kind of impressive to me as well, because it combines recurrence, routing, and narrow-format - all of which adversely impact numeric stability during training.
It beats Qwen on effectively every coding benchmark, which I think are the most relevant ones to HN :)<p>On the other benchmarks it trades blows or beats Qwen; as Thinky points out, looping helps <i>reasoning</i> but doesn't scale memorization, so tasks requiring broad world knowledge without tools/search won't be helped by looping.<p>Still super impressive, and probably replaces Qwen 27B in pretty much any workload!
you can also just make qwen looped / carry forward its reasoning from pass to pass. kind of does the same thing and you dont need to change the weights.
I wonder if one could build a looping model with PLE but with different embedding loaded at different iterations of the same layer.
Beyond that, they detailed quite a bit of their training pipeline for Inkling 2 Small.<p>Apparently trained on a ~couple dozen trillion tokens; which is far smaller than most pretraining and posttraining runs for its size class.
This has been my line of thought for a years. If we are aiming for on device intelligence, looping (or other methods of parameter reuse) are one of the best methods since consumer hardware is primarily memory constrained.<p>Which is counter to the way most open weight models have been leaning. Many current models are MoE (Mixture of Experts) which by their very nature are parameter heavy. They are trading parameter efficiency for compute efficiency which is kind of the wrong tradeoff for current consumer hardware like gaming GPUs.
> since consumer hardware is primarily memory constrained.<p>It depends, people with Macs or AMD with unified memory are more constrained by memory bandwidth than with resident memory.<p>Unfortunately for us, looping doesn't help on that from.
Consumer hardware doesn't really boil down to a single constraint on how much memory is available, in my opinion.<p>Consumer hardware has limited memory, limited memory bandwidth, limited compute, and limited storage. Any model designed for consumer hardware needs to tackle all four, while still optimizing for intelligence.<p>MoEs save on both memory bandwidth and compute compared to a same-size dense model, and importantly, MoEs can still run quickly even with partial offloading. People have even demonstrated getting pretty good performance out of MoEs streaming from SSD with careful optimization.[0] Dense models fall apart if they aren't run entirely in the fastest memory pool.<p>Memory bandwidth is scarce on consumer laptops and smartphones. Looping demands <i>even more</i> memory bandwidth than running the same parameters once, so it is not an obvious win. If looping can make the model significantly more intelligent, it might be worthwhile. It is interesting to see a MoE combined with looping, as with Inkling 2 Mini. On your point about memory being the constraint, looping increases the size of Inkling 2 Mini's KV cache because the looped layers don't share KV between loops.<p>With DSpark-style specdec getting better and better, dense models make more sense than they used to since you can amortize the memory bandwidth better. The problem is then prefill (prompt processing), since specdec does nothing to help with that, and processing long inputs (search results, reading documents, etc) will still feel very slow. A MoE can breeze through that same prefill since it doesn't need as much compute.<p>With the introduction of Engrams/PLEs, model designers now have another axis for adding knowledge to models at the cost of storage, with comparatively little additional compute or memory bandwidth.<p>There are tradeoffs on everything. Despite my own extensive testing of different models on different hardware setups, I have not found any one-size-fits-all model architecture for consumer hardware. It is a hard problem. Small models are getting better all the time, and I think even smartphone-sized models will pass the threshold of usable intelligence for most people in the next year. Spark-X2.6-4B[1] and LFM3-1.7B[2] are surprisingly good for their sizes. Inkling 2 Mini looks great, supposedly offering Qwen3.8-27B's intelligence with much less compute required.<p>[0]: <a href="https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B#on-device-capability-and-efficiency" rel="nofollow">https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B#on-devi...</a><p>[1]: <a href="https://huggingface.co/XHToken/Spark-X2.6-4B" rel="nofollow">https://huggingface.co/XHToken/Spark-X2.6-4B</a><p>[2]: <a href="https://huggingface.co/LiquidAI/LFM3-1.7B" rel="nofollow">https://huggingface.co/LiquidAI/LFM3-1.7B</a>
> A MoE can breeze through that same prefill since it doesn't need as much compute.<p>Important caveat: if not fully offloading[0] then the MoE model's prefill speed completely tanks, as prefill is heavily dependent FLOPS and parallelization and CPUs suck at that compared to GPUs.[1]<p>[0] 'Offload' is being used in the llamacpp sense of using the GPU instead of CPU<p>[1] Generally, prefill is bottlenecked by compute while generation is bottlenecked by memory bandwidth
Fortunately, that's not really how it works anymore.<p>When you're prefilling, you can afford to amortize the cost of sending the experts to a scratch space on the GPU then running a bunch of tokens through that set of experts, then repeating with the next set of experts, and so on. There's no reason to compute those on the CPU, and llama.cpp <i>already does</i> move prefill to the GPU by default for partially-offloaded MoEs.<p>Where performance would still tank is if you don't ensure the microbatch (ubatch) is large enough to amortize that cost, and if you're prefilling only a small input, it's not going to reach full speed. For a large input, you can think of each expert as visiting the GPU once for every ubatch of input tokens. A larger ubatch means fewer transfers.<p>Try increasing `-ub` (2048 is a sane default, but 4096 or higher can sometimes be desirable), and watch the partially-offloaded MoE prefill performance magically go way up on llama.cpp for large inputs.<p>EDIT: Also, if you haven't tried replacing `--n-cpu-moe` with `--cpu-moe` + `--moe-cache-mib` (with all of your GPU memory that you can spare), it is even better than `--n-cpu-moe`, since real world models don't distribute work over the experts in a perfectly random way. The moe-cache speeds up decode/output tokens per second.
> Wonder what it means for safety and interpretability at this level of availability.<p>We already know from official model cards that the GPT-6 models are substantially less monitorable than other (non-looped) models. Which is not surprising, since they are doing some amount of Neuralese thinking between each token.
This loop thing isn't as new as it sounds. There were vision models pre 2023 doing these sorts of things.<p>The safety thing is more or less openais way of saying "be afraid we will be the ones to destroy the world not anthropic". The truth is, all of these models looped or not lack real interpretability. As far as safety goes no LLM is provably safe.<p>So it's cool people are picking up on the first steps to explore to make more efficient models, but it's literally just a start and there is so much more that people can do. It's also nothing to be scared of, llms in general are scary when held wrong(they almost always are).
I am getting Stable Diffusion level of breakthrough vibe here. And such looped transformer SMALL models are best suited to embed directly into silicon/transistor. So you get better than ASIC performance by just looping over the hard-coded silicon. Just like we had a breakthrough with "deep" neural network, soon we may have something similar with "deep" loop over the same set of weights. Imaging transistor level looping 128 times. Extreme power efficient, and god knows what level of intelligence is achievable. I could very well foresee purchasing USB Intelligence Devices with some standardised driver protocol. You just buy one off Amazon, throw away the old one and get an Intelligence Upgrade.
M1 Pro Macbook, 32GB: About 10t/s using q4 xl while running browsers/etc. and interacting using the local browser.<p>test 1: write a python program to sort a csv by lastname and print out fields.
Result: Nailed it. Nice python program. Nice explanation of output, if a little chatty (it gave an example output, which I didn't ask for).<p>test 2: my standard "identify the type and year of this coin" test. Failed miserably, I gave up after 12,000 tokens of it babbling incessantly about possible misinterpretations of a silver half-dollar. Unsurprising - most small models suck at this test - but slightly annoying it seemed to get in a loop.<p>Pretty good performance and result for the (very simple) code generation problem, though. One of the first models that seems OK running side-by-side with a browser on my older MBP.
Did you give it search tools?
Usually a model with access to curl can figure out how to do a search.
No, this was just in the really basic llama web ui
how about qwen 3.8 27b on the same machine?
It doesn’t quite fit 32 GB and is therefore really slow.
On that machine, I get 2–4 tok/s with qwen3.8 27b :(
What reasoning level in test 2?<p>Seemed to default to extra-high on Unsloth for me with Unsloth's Q4. (And yeah, it absolutely overran on a test I did.)
Quite an interesting model, but the second of my test prompts in "extra high" thinking in Unsloth really underscore what they say about knowledge here:<p><i>> While looping improves the model’s reasoning, it doesn’t enable its weights to store more facts—knowledge is mostly a function of total parameter count. Inkling 2 Mini’s score on AA-Omniscience, a measure of factuality, is in line with other models of the same size [...] but knowledge-intensive benchmarks doesn’t improve with more passes...<p>> Inkling 2 Mini overcomes the knowledge gap by efficiently using search and tools.</i><p>I tested the Unsloth Q4 version.<p>My prompt asked it to write a singleton class, single file plugin for WordPress that hooks user login and stores last login time as metadata. It has some code style requirements. The prompt also asked it to make no tool calls so it didn't search and only used its knowledge.<p>The model burned through 19200 thinking tokens on this task (which was the hard limit I had set) and didn't finish. Its thinking suggests it was going to make at least one mistake.<p>Both Qwen 3.6 35B and Gemma 4 26B can cope with this task, from memory, in far fewer tokens (even taking into account Qwen's tendency to loop in thinking).<p>Rerunning the same prompt with tool use allowed, it searched with browser use, pretty effectively, and got to a solution, but it still made the mistake the thinking traces suggested it would earlier (wrong hook name). Other than that it wrote good code.<p>But 20K-ish tokens in, I could then not get it to search to make the required correction — it would repeatedly EOT rather than actually run the tools it was saying it would call. Whether this is the model failing or Unsloth's harness and the model disagreeing, I don't know. Quitting and reloading the model to continue the conversation didn't help.<p>It worked through my usual MySQL-5.0-compatible max score per category prompt nicely, though it didn't volunteer the suggestion of adding an index for speed, but it doesn't outperform LFM 2.5-24B-A2B, which is a non-looping model of similar size. It does work through SQL in its thinking traces quite nicely, though.<p>Car wash problem got a correct answer quite quickly, albeit with a surreal caveat (why do they always seem to need to offer a caveat, no matter how weird)<p>So far, I think I would run Muse Glimmer or Nemotron Bolt over this; Muse Glimmer would output tokens more slowly but still get to a result much more quickly.
The GGUF version of this that can run on Ollama/Llama.cpp is: <a href="https://huggingface.co/unsloth/Inkling-2-Mini-GGUF" rel="nofollow">https://huggingface.co/unsloth/Inkling-2-Mini-GGUF</a>
Recently SensaNova also released a looping model with unbelievable numbers<p><a href="https://www.sensetime.com/en/news/sensenova-loop-9b-release-20261118-1435" rel="nofollow">https://www.sensetime.com/en/news/sensenova-loop-9b-release-...</a>
Not a single comment when it was posted here: <a href="https://news.ycombinator.com/item?id=50531941">https://news.ycombinator.com/item?id=50531941</a>
> You need to agree to share your Hugging Face information to access this model<p>the mixed signals sent by publishing weights to hugging face, but restricting downloads, is just too much for me to untangle.
This looks even more impressive at 9B parameters.<p>Has anyone tried this? Does it run well with the latest llama / unsloth servers?
I'm wondering when will these local models be trained for expert activation on a task-by-task basis, even in an implicit manner, by having a loss function that encourages expert coherence across token activations. The model size could be much, much higher as the bandwidth penalty would have to be paid only occasionally.<p>This would be an advantage that would only benefit local setups, which is why I guess we don't see it yet.
Very, very excited about models of this size. Small enough to run on a Mac Mini!<p>Ran my 43 question data analytics exam on it against OpenRouter & my M4 Pro 64GB Mac Mini via llama. It scored a "B" on the exam [1], same score as DeepSeek V4.1 Flash and Sonnet 5.<p>The main limitation that I saw was a "runaway loop" when asked to find a issues in a dataset: it worked for over 3 hours on my Mac Mini, taking over 180 tool calls to answer the question before I cut it off.<p>It's the most accurate U.S. open weight model that fits on a Mac mini that I've tested (other two are Muse Glimmer 30B and Gemma 4 31B). Against other models its size, the only model that got a better score was Qwen 3.8 27B(it got a B+).<p>On my Mac Mini, each question takes between 4min-30min to answer. On OpenRouter, the median was ~70s. So ~4x slower on desktop software. That should get faster when mlx support comes available.<p>For memory usage, Llama was taking about 29GB on my Mac Mini.<p>Very exciting overall!<p>[1] <a href="https://plotly.com/blog/inkling-2-mini-plotly-data-analytics-bench/" rel="nofollow">https://plotly.com/blog/inkling-2-mini-plotly-data-analytics...</a>
I think this model will be a crazy big deal, it might be the first model that can reliably run multi-agent reasoning harnesses on local hardware, as this is a MoE model and can have several agents each with their own KV caches that fit into VRAM.<p>Dethroning Qwen 3.8 is a big deal too, and I like their section about CoT monitorability. (This model beats Qwen on benchmarks - TML is not known to benchmaxx nearly as hard as Qwen.) I guess TML really are a bunch of rationalists who know what they're doing, sort of like Anthropic.
> TML really are a bunch of rationalists who know what they're doing<p>yes<p>> sort of like Anthropic.<p>no. Anthropic are not "rationalists," they are absolute fear-mongers. from "no software developers in 6 months" to "this could enable bioterrorism so we do not even let you mention the word 'virus' in any context", I do not see a hint of rationality.
I wonder how Ant's corporate culture will change now that so many employees will have secured some quite luxurious nest eggs.<p>Maybe they'll chill out a bit?
I mean I thinkhe meant like big R Rationalist like Effective Altruism or whatever other spin offs is du jour.
I'm gonna bet you'd be complaining a lot more if a bioterrorist used Claude and their blocks weren't strict enough.
Going to test this right away on my AMD 7900XTX and see how it does compared to Qwen 3.8 27B (or DaisyRose 27B that I'm currently running). Given that Inkling 2 Mini is a MoE, it should provide way higher decode speed, but with comparable performance. Exiting!
>Exiting!<p>OK, but please come back and let us know your results!
Got it to run, but a few bumps in the road – mostly because this is a very fresh model.<p>Flash Attention (FA) is not supported on the current llama.cpp (at least not for ROCm) for this model so KV cache is forced to f16 for the value (V) and I cannot lower it to q4 go get decent KV cache. But the key (K) could be set to q4, and that got me about 100k of ctx. Current benching shows ~28 tok/s decode and ~481 tok/s prefill.<p>It's a start, but I expected better speeds even on a non ideal setup. I'll keep experimenting!
Hey, just in case.. I have been maintaining this fork of llama.cpp, I run also 7900XTX, 3 in my case... and I want to know if you see improved speed... notice the ngram mod where instead of loosing the ngrams they are stored in disk, I have a 31M slot and so far it's being working really nice for me<p>I'm seing around 2800t/s on pp at the very beginning of the conversation, and then going to around 800 by the 256k context size...<p>I've also update the value in the gguf to have instead of context size 256k be triple and have use it all the way to around 600k context size<p>use the cache eviction branch
<a href="https://github.com/alainnothere/llama.cpp/tree/disk-cache-eviction" rel="nofollow">https://github.com/alainnothere/llama.cpp/tree/disk-cache-ev...</a><p>llama-server
-m /path_to_model/Qwen3.8-27B-UD-IQ4XS_u.gguf
-np 2 --host 0.0.0.0 --port 8089 --kv-unified --temp 0.2 --top-k 20 --min-p 0.0
--cache-disk-path /path_to_store_conversations_to_disk
--cache-ram 100 --cpu-range 0-23 --reasoning-budget 4096 --reasoning-budget-message "... I am thinking for too long -- let me gather more info about the task."
--no-cache-idle-slots
--spec-draft-auto --spec-draft-sampling probabilistic
--spec-draft-n-min 1 --spec-draft-p-min 0 --spec-ngram-auto
--spec-ngram-mod-size 35544320
--spec-ngram-mod-cache /path_to_storage_of_ngram_mod_this_is_specific_per_model/ngram-mod-qwen3.8-27b.bin
--fit on
--device Vulkan3,Vulkan2,Vulkan1
--fit-target 1024,2560,2560
--ctx-checkpoints 25 --checkpoint-spill-max 25
--cache-type-k f16 --cache-type-v f16 -ctkd q8_0 -ctvd q8_0 -devd Vulkan3
--spec-type draft-mtp,ngram-mod --spec-draft-n-max 7 --spec-ngram-mod-n-min 7
--spec-ngram-draft-n-max 31 --spec-rs-depth 24
--fit-verify on<p>Use this if you have 3 cards, if you have less maybe just remove it
-b 8192<p>In case you want to know how many tokens you have use,
--token-ledger /path_to_where_you_want_to_store_stats/token_ledger.json<p>here is how logs look like I saw a 350t/s on tg but was because you hit ngram mod tables over things the model has already seen<p>457.04.985.808 I srv print_line: token ledger: requests = 27077, prompt = 3928887769 (cached = 3874707808, processed = 47151296), generated = 22126308 (reasoning = 15129037), t>
457.04.985.913 I slot rq_write_fin: id 0 | task -1 | request timeline: task 77431, ttft 1580.0 ms (http 20.4, tokenize 65.1, queue 247.1, select 0.0, restore 2.7, prompt 1244.3),>
457.05.202.640 I srv handle_compl: task created: conversation_id=64d89c6d7ff76923 (tokens 0..18489, system+first_user)
457.05.202.757 I slot get_availabl: id 0 | task -1 | selected slot by LCP similarity, f_sim_best = 0.994 (> 0.100 thold), f_keep = 1.000
457.05.206.552 I slot launch_slot_: id 0 | task 77458 | processing task, is_child = 0
457.05.207.188 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81032..81036 of 81552
457.05.246.405 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81036..81548 of 81552
457.05.310.149 I srv lazify_check: checkpoint: converted pos_min=80943 to disk-backed ('cp_64d89c6d7ff76923_80943.bin', 149.6 MiB released)
457.05.872.722 I slot operator(): id 0 | task 77458 | prompt processing batch, tokens 81548..81552 of 81552
457.05.943.921 I srv lazify_check: checkpoint: converted pos_min=81035 to disk-backed ('cp_64d89c6d7ff76923_81035.bin', 149.6 MiB released)
457.09.000.990 I slot print_timing: id 0 | task 77458 | n_gen = 340, tg = 112.88 t/s, tg_3s = 113.20 t/s
457.11.161.223 I slot print_timing: id 0 | task 77458 | prompt eval time = 790.96 ms / 520 tokens ( 1.52 ms per token, 657.43 tokens per second)
457.11.161.226 I slot print_timing: id 0 | task 77458 | eval time = 5163.27 ms / 642 tokens ( 8.06 ms per token, 124.15 tokens per second)
457.11.161.226 I slot print_timing: id 0 | task 77458 | total time = 5954.23 ms / 1162 tokens
457.11.161.227 I slot print_timing: id 0 | task 77458 | graphs reused = 72319 (pp skipped 525)
457.11.161.228 I slot print_timing: id 0 | task 77458 | graphs prebuilt = 7295 (hit 6975, miss 320)
457.11.161.229 I slot print_timing: id 0 | task 77458 | early prebuilt = 28686 (fit 44246, hit 26411, miss 2275, aborted 0, 96584.2 ms host)
457.11.161.238 I slot print_timing: id 0 | task 77458 | draft acceptance = 0.83204 ( 535 accepted / 643 generated), mean len = 6.00
457.11.161.243 I slot print_timing: id 0 | task 77458 | draft auto = n=6 | 1:55.4 t/s [2.00 tok / 36.1 ms] (4) 2:67.7 t/s [2.50 tok / 37.0 ms] (4) 3:101.0 t/s [4.00 tok / 39.6 ms>
457.11.161.251 I slot print_timing: id 0 | task 77458 | draft by impl: draft-mtp | steps 101 | built 5.5 | offered 5.5 | accepted 4.9 (0.90) | full 80 | step 45.2 ms | 131.5 t/s >
457.11.161.257 I slot print_timing: id 0 | task 77458 | draft by impl: ngram-mod | steps 6 | built 64.0 | offered 14.3 | accepted 6.0 (0.42) | full 1 | step 90.8 ms | 77.1 t/s | >
457.11.161.258 I slot print_timing: id 0 | task 77458 | ngram policy: auto on | cap now 12 | cap min 7 median 12 max 31 over 6 chains | cap bound 6 | whole ema 0.17 | resume 0 | >
457.11.161.259 I slot print_timing: id 0 | task 77458 | spec rollback: prebatch restores 1 | host checkpoints 0 | checkpoint restores 0
457.11.161.260 I slot print_timing: id 0 | task 77458 | chain truth: ngram | chains 6 | censored 0 | would land mean 7.0 | full 0 | hist 0:0 1-4:1 5-8:3 9-16:2 17-32:0 33-48:0 49>
457.11.161.261 I slot print_timing: id 0 | task 77458 | ngram gate: held 38 | agree 5 (would land 0:0 1-4:1 5-16:4 17+:0, pending 0) | disagree 33 (would land 0:33 1-4:0 5-16:0 1>
457.11.161.261 I slot print_timing: id 0 | task 77458 | ext shadow: mtp+ngram | attempts 101 | hits 85 | tested 68 | landed mean 0.4 (over tested) | hist 0:65 1-4:1 5-8:0 9-16:2 >
457.11.161.262 I slot print_timing: id 0 | task 77458 | alt table: n=48 | attempts 101 | hits 45 | would land mean 0.0 | hist 0:45 1-4:0 5-8:0 9-16:0 17-32:0 33-48:0 49-64:0 65+:>
457.11.161.263 I slot print_timing: id 0 | task 77458 | mtp ahead: tried 53 (0 after a p_min stop) | verify full 49 | bonus matched 48 (0 after p_min) | too short 0 | served 48 s>
457.11.161.265 I slot print_timing: id 0 | task 77458 | mtp confidence: stop cap:101(acc 4.9) pmin:0(acc 0.0) other:0 | p bins landed/tested: <0.5 1/3 0.5 0/1 0.6 1/1 0.7 0/0 0.8>
457.11.163.234 I slot release: id 0 | task 77458 | stop processing: n_tokens = 82195, truncated = 0
457.11.163.426 I srv print_line: token ledger: requests = 27078, prompt = 3928969321 (cached = 3874788840, processed = 47151816), generated = 22126950 (reasoning = 15129071), t>
457.11.163.526 I slot rq_write_fin: id 0 | task -1 | request timeline: task 77458, ttft 884.3 ms (http 21.2, tokenize 68.3, queue 0.1, select 0.0, restore 3.3, prompt 791.0), gen>
457.11.396.231 I srv handle_compl: task created: conversation_id=64d89c6d7ff76923 (tokens 0..18489, system+first_user)
Please reply when you have tested it I’m curious.<p>Also do you know what the best local setup is?
I had astra patch in both required llama.cpp PRs and it's passing a few tests locally on my 7900XTX so far, but ~500 tok/s prefill and ~26 tok/s decode are sobering: The new three-time-looping might be cutting speeds. Maybe there's something here still though, maybe it needs less reasoning output for results to be usable/good? More testing needed!<p>edit: found some oudated mesa/radv/shaderc/glslang packages to update, and up to 34 tok/s now. MORE MORE<p>edit2: trying to see if this dspark drafter is usable and will give a good speed boost, didn't load up on first attempt. I'm probably too early [5-10min later] ... found an unsloth llama.cpp build with vulkan dspark support and it actually slowed down decode :(
I'm getting 700 tok/s prefill and 52tok/s generation on a 5090 RTX, not bad
That seems slow, should be much much faster hmmm
It's bad! I got 150 tok/s on RTX 3090.
5090 RTX is like 1800GB/s? I bet you were seeing much higher speeds on Qwen 3.8 27B right?
Hm, that seems strangely slow. Qwen3.8-27B/DaisyRose-27B Q4 currently run with ~66 t/s and ~800 t/s prefill with 3 slots and about 190k kv cache. It performs really well so far.
I'm always on the lookout for reports from my 3090 friends<p>Looking a bit oof on reddit rn, I'm seeing reports of 10 tok/s via llama.cpp :|<p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1y1g9cr/inkling_2_llamacpp_support/" rel="nofollow">https://www.reddit.com/r/LocalLLaMA/comments/1y1g9cr/inkling...</a><p>Edit: Hopeful that this gets the ninfer treatment, if the benchmarks are to be believed it feels like there is probably both amazing speed and real performance to be unlocked by the community here on the x090 cards
For 7900XTX you mean? I run Fedora Silverblue 44 at the moment and could share my current containerized setup with DaisyRose 27B.
I don’t understand the business model of these labs which obviously spend a huge amount of money on compute then release their models. I definitely glad they do, but how does it make business sense?
TM the company also sells FTaaS (fine tuning) with a very streamlined ux/dx. This model came at a time where "small" models are in the 200+B params, and many "local" folks can't run those on commodity hw. By being so small that you can run this on almost anything local could lead to an increased demand in their FT service (and I hope it works for them, so they're incentivised to continue releasing true small models).<p>There's also the aspect of b2b where someone could go to TM with their own data, and have them assist with finetuning and deliver a model streamlined for their own flows / tasks. Sorta like what mistral are doing in eu.
I think we will see a lot of small labs with funding. The funding amounts may be small but it is no secret that labs from around the world want to build usable (tiny/small/large) language models. Even if investors are not sure how or when the returns will come, they want to see traction and usage. Enough people have enough money to spend. It may sound odd but I think this is the largest spend of private wealth in research that we have seen globally. I see small investor willingness all over the place and I am far away from tech capitals.
Kinda the same (directionally if not in magnitude) as companies that open source their core product (thus giving away millions of dollars in dev salaries, cumulatively up until release and also continuing for maintenance) but do managed deployments and consulting to pay the bills.<p>So, it’s advertising? For customers but also for any big companies who want to acquihire a really competent team and have them do the thing but in-house.
Traditionally governments invest in basic research because that's how they discover leads that enable civilizational jumps in technology. If you neglect particle physics research in the early 20th century you face existential threats to your sovereignty in the 1950s.<p>I can see large investors thinking in these terms - make lots of relatively cheap research bets to try to find the leads to tech that might make similar technological leaps within 15 years.
Starter AI labs do it for clout so they can hire the best researchers, raise more money, and then do real frontier training runs.<p>Presumably Thinky will do the same here if they want to win.<p>Relevant quote (but relating to the Chinese labs) from twitter, saying this better than me [1]:<p>"most Chinese labs do open sourcing for a host of relatively mundane reasons (mindshare capture beyond their inference capacity, prestige and publication opportunities to lure in young researchers, keeping up with the Joneses, the decaying factor of harvesting community technical support and post-training effort, the increasingly uncertain factor of extra revenue via hosting license terms, some mild marginal bonus to state favor)"<p>[1] <a href="https://nitter.r4fo.com/teortaxesTex/status/2121034214556754349" rel="nofollow">https://nitter.r4fo.com/teortaxesTex/status/2121034214556754...</a> Actually a very insightful comments on DeepSeek under Wenfang.
it’s structured as a PBC. getting return on investments is the investor’s headache, not necessarily the company’s. we don’t know what deals have been struck.<p>edit: they also have this <a href="https://thinkingmachines.ai/tinker/" rel="nofollow">https://thinkingmachines.ai/tinker/</a>
I’m sure there’s tons of speculative money to be made via acquisition.
Same model as companies like red hat. They hope to sign enterprise support deals for either managed services or bespoke updates
For Inkling in particular, this looks like one of their potential revenue sources: <a href="https://thinkingmachines.ai/tinker/pricing/" rel="nofollow">https://thinkingmachines.ai/tinker/pricing/</a>
This looks great and is getting a lot of fun.<p>Interesting cyberbench: <a href="https://x.com/ValsAI/status/2123874523938497007" rel="nofollow">https://x.com/ValsAI/status/2123874523938497007</a><p>Thinking machines just got up in Evaluation.
Models like this are a product of resource constraints. It's great. It's why I am for a data center moratorium. Scarcity breeds innovation. If we're forced to work with what we have, incredible efficiencies are on the horizon.
when considering gpu hours, larger models are easier/cheaper to train than smaller models. maybe if datacenter inference is infeasible, then it will convince model providers to train small models, but from a pure training perspective, smaller data centers lead to larger models.
What are some models in this ballpark that are also good at conversation? They always show off their software capabilities, but can it talk naturally? make creative stories?
The closest to this that's also good at prose is gemma4 26b a4b. That's what I'm using for a fully-local "Tunnel Goons" (light D&D-style game) harness where multiple player agents and a DM agent improvise a story together.<p>I haven't tried inkling so I have no idea how it speaks, writes, and tells stories. I'd love to know what other small models folks are having success with in role-play, gaming, creative writing environments.
Yes — Gemma writes in the most natural, least-annoying way.<p>The 12B model is super too; I think it might be a better writer than the 26B.<p>Getting it to write about uploaded photographs and drawings is modestly entertaining and perhaps useful, though I haven't really found a use that is worth checking in on too regularly. It did point out some things about my own photos that were possibly valuable, though.
Yeah, I've settled for gemma 4 too, but I'm always in search of a better model who <i>gets</i> what I'm asking
Very cool, and the also released a research version with <i>variable</i> looping support: <a href="https://huggingface.co/thinkingmachines/TML-d24-varloop-research" rel="nofollow">https://huggingface.co/thinkingmachines/TML-d24-varloop-rese...</a><p>Idea being you can configure at run time how much thinking you let the model do per token.
The demo video was 32 tok/s and did a 15X speedup when generating the code? Weird that they didn't publish token speeds for this laptop setup they recorded. I'm worried this tri-loop setup is going to be too slow. QFN is impressive but lately been feeling a little brain-fried by its prose.
Big fan of open weights but those Terminal Bench numbers are pretty abysmal, and I find that tends to track "agent usefulness" more than most SWE benchmarks.
Benchmarks look good, and inkling 2 small felt better than the benchmarks suggested, so I am very excited to play with this.<p>From a research perspective, I'm even more interested in the varlooped variant they're releasing <a href="https://thinkingmachines.ai/model-card/tml-d24-varloop-research/" rel="nofollow">https://thinkingmachines.ai/model-card/tml-d24-varloop-resea...</a>
I'm not aware of any other working open source models like that.
My current feedback on Inkling 2 Small: 1. it feels intelligent, and can solve open problems throw at it like other quasi-frontier models (DeepSeek 4 Flash 0731 / DeepSeek 4.1 Flash 1017 / GLM 5.3 Flash etc). 2. its responses sometimes can be odd (they are honest about it in the blog post!), can have wrong use of markdown syntax, sometimes overly brief, etc. 3. tool uses are good, but sometimes have some old take on tools (too quickly to assume image generation tools cannot do text etc).<p>EDIT: by "open problems", I meant the problems I don't really formulate how to solve it first in my mind and then throw to it. I just describe what I want and let it to figure it out.
Sounds like <a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16" rel="nofollow">https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4...</a>
Very interesting to see them publish a looped transformer. It's been reported that OpenAI are now using looped transformers, though of course not confirmed.<p>I think this may be the first public example of a looped transformer model good enough for real world use cases?
very promising with benchmark, I have used Inkling 2 small earlier and it was good with vision in the loop
The thing I like about looped transformers is it implies theres some kind of modularity learned with how each iteration works on the residual stream, which sounds like a lovely inductive bias.<p>(all vibes, no data from me)
I tested the regular Inkling 2 and in my tasks it’s worse and more expensive than Hy4-flash. Those are ai agents tasks though so not sure how these benchmarks are scoring so differently
Seems like this is the successor to Qwen 3.6 35B that we've been waiting for? I'm not aware of any other recent models of this size that can run well on mid-tier GPUs without severe quantization.
Pelicans aren't terrible (for the model size): <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fb264d7524e46f1a2d1bf78ad5441e2c4" rel="nofollow">https://tools.simonwillison.net/markdown-svg-renderer?url=ht...</a><p>Generated with Inkling-2-Mini-IQ4_NL.gguf and llama.cpp on a M5 MacBook Pro.
bonkers - I'm getting 20 tokens/s running the 4-bit version on my 2021 M1 Max MacBook pro with 64Gb...!!<p>recipe: <a href="https://gist.github.com/gregsadetsky/85e8b8ea91430dc826611b97e5efa27c" rel="nofollow">https://gist.github.com/gregsadetsky/85e8b8ea91430dc826611b9...</a>
I'm getting about 18 on the same hardware using the unsloth build of llama.cpp; I guess I should try upstream now that support is merged.<p>Theoretically with 400GB/s memory bandwidth and effectively 5.4B active parameters (accounting for the looping), token generation should be closer to 74 tokens/second. 18tps is close to what you'd expect for a dense model this size (22GB for UD-Q8_K_XL). Hopefully over the next few days, someone clever figures out what's going on and optimizes this.
> Theoretically with 400GB/s memory bandwidth<p>Well it would be theoretical, alas — the M1 Max can't actually saturate the bus with the GPU. I have seen varying suggestions that it might in practice be anywhere between 320GB/s and 390GB/s; for the CPU it is much much lower (only 243 GB/s).<p>I am seeing 16-20 tokens per second generation in Unsloth Desktop. And if they can get to higher speeds it might indeed be interesting.<p>But where you say:<p>> and effectively 5.4B active parameters (accounting for the looping)<p>It says 5.4B "weight activations per token" — but it is a 3.4B-active model, so my interpretation of that is: it's activating a total extent of 5.4B <i>different</i> weights, but is it not activating a different 3.4B subset of this 5.4, three times in succession?<p>So your calculation should be based on 10.2, perhaps? (Not taking into account that some layers might not loop)<p>(These are extremely late-night thoughts, I should be in bed and not on HN, so I am probably wrong)<p>ETA: just found a discussion of this here, relevant to this particular model, that I am going to read tomorrow:<p><a href="https://www.reddit.com/r/LocalLLaMA/comments/1y1ip5f/discussion_in_the_age_of_loops_what_do_active/" rel="nofollow">https://www.reddit.com/r/LocalLLaMA/comments/1y1ip5f/discuss...</a>
> a looped transformer, running the middle layers of the model three times for every token.<p>Completely unqualified take here, but unconditional re-runs of a fixed set of layers doesn't strike me as meaningfully 'loopy'. The exact performance profile is available in a more trad architecture by just inserting the layers after one another?<p>My presumption on loopy architectures was that they looped longer on token selections that were more difficult / higher stakes, with automatic kick-outs after some <i>n</i> loops to avoid machine analysis paralysis.<p>ALTHOUGH I guess that the presented architecture may also be useful insofar as it can fit in less memory than the same model with a tripled set of , and possibly that training over a looping architecture converges faster / truer than over naive architectures of the same size.
excited to see Thinking Machines accelerating their effort, but in their own way. loving this almost more indie flavor of AI research
IIRC Macs with 32GB unified memory have a default soft-cap of ⅔ for how much memory the graphics subsystem can use, so when they say it's targeted at 32GB Macs that effectively means targeting 21.3 GB VRAM?
I routinely override that cap, kill web browsers, etc., and shoe-horn large models on my 64G Mac.<p>EDIT: for 64G Mac, I use: sudo sysctl iogpu.wired_limit_mb=54000
I've been thinking about how small looped models are probably going to be the next big thing in local LLMs, and here it is. Exchanging a lower memory capacity requirement for more compute instead.<p>If the GPT-6.x models are looped, no doubt memory was a big reason.
Inkling's benchmarks for testing their model are really amazing. There is no better way to demonstrate a model's capability than building a complete game/application powered by the model itself. For eg: the tiny pixel town where all the 20 villagers are powered by the model on a MacBook - is rock solid!
This looks so good. Might replace Muse Glimmer for me for local tool calling workloads. As good as Qwen 27b for many tasks and no state mandated censorship - yes please! Super excited to try this out today.
This is pretty good but also from their hf model release;<p>>> Note: each extra pass through the looped layers adds about 1.0B effective active parameters. Since the model runs with three passes, there are 5.4B effective active parameters per token.
> 3.4B active<p>Does this mean it can run on a GPU with just 4GB of memory?
All this means is that only a fraction of the model is read per token, so it's no big deal if the parameters are resident in system RAM.<p>Precious VRAM is still required to store the KV cache, the expert router and whatever dense parts of the model which are there. What's left is used as cache for experts.<p>Additionally tools such as Strata juggle the experts between VRAM, RAM and disk.
Wow, that coding demo was super impressive. Can't wait to try it out. Just need to wait for the hardware market to cool down a bit so we can run these on our own commodity hardware.
The masses lying in wait for Qwen 4 cry out in joy. A savior? Could it be?!
When you are sure enough about what you are doing to include Qwen 3.8 27B in your compare.. we know you are talking... let's test drive it!
I'm guessing one of these local model companies will get picked by Apple sooner or later. It is just so within their business model and they have the hardware.
Probably not for the foreseeable future. Apple penned a "multi-year collaboration" with Google to build their foundation models for them: <a href="https://blog.google/company-news/inside-google/company-announcements/joint-statement-google-apple/" rel="nofollow">https://blog.google/company-news/inside-google/company-annou...</a>
Apple refuses to pay ai engineers an ai engineer salary. They're not going to acquire one of these companies until the cost of issuing them is within their comfort zone.
In my tests inkling-2-mini shows results similar to GPT 5.6 with light reasoning. Impressive ...
I didn't even know it was a thing to train using nvfp4 4-bit floats - I always thought it was a post-training quantisation thing. Fascinating!
I've been experimenting with local agent committees, and this should allow me to get a couple of agents running in "realtime" instead of sequentially.<p>It's funny, you think you have enough RAM then you realize that what you need is more memory bandwidth (ugh).
I tried warning Apple Silicon LLM adopters. You can buy a $1,000 GPU with memory bandwidth that is comparable to a $30,000 M5 Ultra Mac Studio (and likely faster at GPU compute too).<p>If looped transformers on smaller LLMs is the future, then the VRAMlets win the war and the Unified Memory crowd overpaid for DDR5.
Looking at the link to quantizations it looks like unsloth has already quantized it:<p><a href="https://huggingface.co/thinkingmachines/Inkling-2-Mini" rel="nofollow">https://huggingface.co/thinkingmachines/Inkling-2-Mini</a><p><a href="https://huggingface.co/unsloth/Inkling-2-Mini-GGUF" rel="nofollow">https://huggingface.co/unsloth/Inkling-2-Mini-GGUF</a><p>In terms of the active parameters and token/s on slow or less capable hardware it seems like it should be in the same general performance category as qwen 3.6 35B A3B.
What coding harness would someone use with this, such as in the article’s demo video? The demo was impressive
"Inkling" is strongly associated with Nintendo's Splatoon series these days, FYI.
Omg I've been working on the same thing! I was doing this off a hunch from David Noel ng's RYS posts!<p><a href="https://dnhkng.github.io/posts/rys/" rel="nofollow">https://dnhkng.github.io/posts/rys/</a>
I tried that for awhile too. Fun tip: You don't have to retrain the entire model after looping it to make it coherent again.. Just pick layers carefully then repair the residual by doing some parent distillation. ;)<p>That took me from needing to rent space in avast to just running it locally on my sparks and getting comparable results.<p>I also discovered that you can fix a lot of the damage caused by quantization the same way. Just do 200-500 rounds of parent student distillation with everything frozen but the residual, often the quantized version will improve drastically.
When inkling first released I tested it out and thought it was the best prose from an LLM in my quick sampling -- easy-reading and easy to understand. Big hopes I can write code with one of these models one day.
Nice model, i get 17 tokens/s on my AMD RX 9060 XT 16 GB. Testing coding with it now.
Thank you! Added to the Tiny League (<a href="https://tinyleague.nunodonato.com" rel="nofollow">https://tinyleague.nunodonato.com</a>)
This is actually a big deal. I think probably by 2028 we will get a large number of daily tasks done reliably by laptop size local models.
What is the marketing pitch and the one-liner for what makes this model special?
Even more talkative than Qwen
hate these dark practices. they say click here to try it on our site. then u click. auth wall. fine you give it all your info. sunk cost is maximized. then they hit you with the credit card wall.
Finally another model for local LLM users.<p>To achieve better performance, compared to parameter scaling methods like mixture of more experts and more whatever embeddings which throw parameter efficiency out of window, layer reuse or looping show higher parameter efficiency and is much appreciated, which I think is enabled by attention residual?<p>The problem I observed is that in SWE workflows the output length of inkling is longer than qwen 3.8 27b xhigh, which is already notorious. Multilingual support except English is another pain point.<p>Anyway, congratulations to the inkling team for this release.
Got between 130 and 152 tokens per second for Q4 using ninfer on a RTX 3090. Pretty good.
I'm starting to believe Gestalt is actually theirs. This is very nice.
Yikes, the pelican was insane
Sensanova 9b is another one
So the transformer launched in 2017 is still the foundation of these current models
Doesn't seem to run in MLX-Serve yet :(
This looks like it could run on my M2 32GB MBP.<p>I honestly would rather have nice, focused-purpose local models than huge "let's make AGI" datacenter-hosted models. Will have to try it tonight when I get home.
I'm loving this trend. My framework desktop is looking like a better and better investment.
Indeed. I 've been using Sage A1.5 35b-A3b to have more speed because of strix halo limited memory bandwidth. Looking forward to testing Inkling-2-mini once it has lamma.cpp support. Genuinely excited from a model outside Qwen3.6-35B-A3B ecosystem competing in that space.<p>P.S. Qwen3.8-Flash-Next is also a very good model for Srix-Halo/128Gb Framework Desktop but I liked to have more RAM for other things as well even when I was running it in Q4. Usable though.
What are you running on it? DSF 4.1?
tell me more about this framework desktop..
Surprised they used a log scale for tokens per second. Seems like a missed opportunity to nail home the difference visually.
Which inference engine works right now? <a href="https://github.com/ggml-org/llama.cpp/pull/32694" rel="nofollow">https://github.com/ggml-org/llama.cpp/pull/32694</a> seems to have some comments saying it doesn't work well. Which one gives more than 200k context on 4090 for a Q4? Currently Qwen3.8-27B-Q4_K_M gives me that much space.<p>Also, why the model file size is only 20GB while UD quants are much larger for Q_8?
pretty cool I guess I just don't know what i'd use it for other than being a nice toy. if it can fit on smaller embedded devices that could be useful
Can anyone running one of these locally give any sense of how good they are?
[dead]
Somewhat astonished that thinking machines would do this considering that looping transformers a.k.a. neuralese is safety wise way harder to keep track of what the LLM is actually doing. by the way, thinking machines CEO is one of the people that quit openAi over safety concerns.
> Efficient intelligence makes Inkling 2 Mini a great choice for inference where speed is paramount, or for running on local hardware.<p>Presumably v. good for offensive use cases too.<p>And a looped model?<p>This and GRH is +0.1% on my p-doom today.<p>The trends of capabilities and efficiency we've seen over the last few weeks should terrify anyone with any sense.<p>This will not end well but we'll have some nice AI generated dashboards until then I guess.
Anyone following the news is aware that we've had Mythos-comparable cyber red-teaming capabilities for quite a few months now. The sky hasn't fallen, instead people have been using it for reverse engineering, postmortems, pentesting and decomp projects. In a world where frontier labs and governments can coordinate large-scale cyberattacks with AI, having an uncensored open cyber model is necessary for defending against them. The alternative is much worse.<p>But hey, you're welcome to increment your P(doom) until the year 2057 if it makes you feel better. Maybe the government will finally put their foot down with the release of Qwen 28, which according to OpenAnthropicTimeWarnerPepsico will have unprecedented puppy-kicking use cases that necessitate regulation.
We've been getting hit pretty hard in Japan with cyber attacks. It has been very noticeable. So, I respectfully disagree a little.
You may not be fully seeing the trends are and over indexing on the now. You may also have different views on instrumental convergence to me which means that you don't see the same exponential threats as I do as AI continues to improve beyond this point. You may also have different views on how likely the power and capabilities AI will provide humans will be leveraged and how easy it will be to defend against those with bad intents.<p>I can't say you're wrong so won't be snarky with you. I think your view is legitimate and broadly held. Maybe safety and defensive techniques will scale with capabilities (both model capabilities and resulting technologies of AI). I see several sets of paths which would be likely to end in a doom scenario for humanity. A world in which exists a god-like intelligence, which we somehow retain control over, yet don't use it to concentrate power or harm other humans seems very unlikely to me.<p>But I hope so much that I'm wrong that I genuinely appreciate anyone who takes the time to tell me how much of a moron I am for worrying as much as I do =)
I think you are making hysterical extrapolations. Your original claim is that recent discoveries like looping transformers and theorem progress is inching humanity closer to doom. But this is part-and-parcel with scientific advances that have happened <i>before</i> AI, and a clean continuation of those efforts. Preexisting power structures (like mutually assured destruction, neoliberalism and the military industrial complex) will not disappear even if we get a god-like AI. Your concerns could be recycled to criticize <i>legitimately</i> threatening scientific progress being made by American/European adversaries in a military sense, and it would be a lot more immediately concerning than praying to God that we don't get an Omnissiah from 40K. One is realpolitik, the other is crank philosophy.<p>If you can delineate an actual threat towards humanity that frontier AI possesses, I'd like to hear it. Cyberattacks <i>are</i> a valid concern, but we're going to see AI-assisted cyberattacks <i>no matter what</i> happens with open AI releases. This has been obvious ever since the HuggingFace attack that OpenAI irresponsibly carried out; we need defenders to have parity with government-sized orgs and frontier labs. If we don't have these options, we've created a power imbalance that favors autocracy and punishes free speech, populist content and blue-teams. The red-teaming aspect is much less scary (to me) than the possibility of governments and labs monopolizing cyber capabilities, which is a real and preeminent concern with proposed regulation of AI.<p>Maybe I am blind - I really am fully open to having my mind changed here. What I'm <i>not</i> willing to do is roll over and accept vague suggestions that the sky is falling and proposing draconian controls on technology that threatens OpenAI/Anthropic's bottom line. They are not trustworthy orgs, and I want my local AI models to be as capable as possible to hedge my bets against their boneheaded, hyperbolic and unaccountable "safety" researchers.
May I suggest adding Comcast to that name
[flagged]