Hy3

(hy.tencent.com)

484 points by andai19 hours ago

24 comments

  • simonw18 hours ago
    Pelican from a few days ago: <a href="https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;6&#x2F;hy3&#x2F;" rel="nofollow">https:&#x2F;&#x2F;simonwillison.net&#x2F;2026&#x2F;Jul&#x2F;6&#x2F;hy3&#x2F;</a> - I was using the free tier on OpenRouter, which expires on July 21st.<p>I tried the preview model 41 days ago and got a pelican with a &quot;change pelican color&quot; button: <a href="https:&#x2F;&#x2F;static.simonwillison.net&#x2F;static&#x2F;2026&#x2F;hy3-preview-pelican.html" rel="nofollow">https:&#x2F;&#x2F;static.simonwillison.net&#x2F;static&#x2F;2026&#x2F;hy3-preview-pel...</a>
    • tarruda17 hours ago
      Recently tried the pelican test on GPT-OSS which was probably one of the best local models of 2025. So cool to see how models have improved in the SVG pelican!
      • pelagicAustral2 hours ago
        I&#x27;m skeptical, as tests go, I think that&#x27;s burned out now. They could easily be training specifically to get a better pelican...
        • gbalduzzi28 minutes ago
          If they were training specifically for it the result would be much better
    • ignoramous13 hours ago
      Curious why TFA calls out &quot;Tencent in China&quot;.<p><pre><code> tencent&#x2F;Hy3. New Apache 2.0 licensed model from Tencent in China </code></pre> Is there a Tencent AI lab elsewhere (MiniMax have some association with Tencent, for example)?
      • JeremyHerrman7 hours ago
        I think it&#x27;s just their version of &quot;Designed by Apple in California&quot;
      • nine_k7 hours ago
        Tencent seems to have subsidiaries: <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Tencent#Subsidiaries" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Tencent#Subsidiaries</a><p>Also they have large European &#x2F; South African shareholders.
    • righthand17 hours ago
      [flagged]
      • invictati9 hours ago
        Some things are better left unsaid.
      • kordlessagain9 hours ago
        I will vouch for Simon. I do not know him personally. To me his posts are honest and inspirational.<p>&gt; I have been overly critical and arguing in bad faith about your writing in the past<p>I think you are just critical without a stated valid reason. Arguing in bad faith seems to be a thing if HN history is the judge.<p>And this is coming from a critical thinker, who is a bit tired of people firing off &quot;human slop&quot; comments. The Internet is full of a lot of people, but even when a few are bad apples, it spoils the lot. Maybe that is the intent. Maybe you are just grumpy for your life&#x27;s situation.<p>It is 100% possible to build software entirely with AI. If you don&#x27;t do that, that&#x27;s great! I still code by hand from time to time, and I&#x27;m reading a lot of Rust nowadays, and learning the ropes. I come from a strong Python and Javascript background, plus networking and operations, which I&#x27;m a whiz at. I don&#x27;t do it anymore, but I know how to inform it is done properly.<p>With this power, I can build things nobody wants to build, but me. Doesn&#x27;t mean one has to put it into production, or it has to pass some security test, although with me driving it probably will. It only need be what is important to the end user, the prompter, to matter.<p>I think Simon helps people with this mindset be better at what they love to do. And for that, we should all be grateful.
    • 301275616 hours ago
      Tony the Tiger on a bicycle selling addictive items to children is not new.
  • minimaxir18 hours ago
    A month ago I wrote a blog post about how Hy3 was topping the OpenRouter rankings despite no one talking about it: <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=48317294">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=48317294</a><p>As of today, it has fallen to 8&#x2F;9th on the rankings. I don&#x27;t see a reason where you would use this model over competitors. However, price economics are bit confusing, as currently the effective input price of Hy3 via OpenRouter is now the same as DeepSeek-hosted DeepSeek Flash V4.<p><a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;tencent&#x2F;hy3-preview" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;tencent&#x2F;hy3-preview</a><p><a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;deepseek&#x2F;deepseek-v4-flash" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;deepseek&#x2F;deepseek-v4-flash</a>
    • Miner49er18 hours ago
      I had to stop using it because I was getting rate limited like crazy. Probably why it has dropped. Seemed like they couldn&#x27;t keep up with demand.
    • jubilanti9 hours ago
      Because it was free with generous limits and high availability, until it wasn&#x27;t.
    • andai16 hours ago
      That was the preview model right? This one appears to be significantly better.<p>I mean it&#x27;s still a small model, but at least the benchmark scores (incl. on DeepSWE) went up significantly.<p>It costs as much as Flash, but the benchmarks are on par with Pro (or above in some cases).<p>Of course, benchmarks are mostly meaningless -- the only real benchmark is the actual work you give it :)
    • terhechte16 hours ago
      It is really slow on openrouter and I ran into many http errors.
    • BoorishBears16 hours ago
      Writes pretty engaging prose, finetunes well, now MIT licensed... what&#x27;s not to like?<p>Oh and very good world knowledge for the size: better than than DS4 Flash
      • Azantys14 hours ago
        Do people really use 100B+ models for writing? I am no writer but to me it seems like writing is one of the easiest tasks with barely any logic or reasoning and as long as its not longer than a handful of pages I expect even 8B models to perform great.
        • orbital-decay8 hours ago
          It&#x27;s pretty clear you&#x27;ve never experimented with it. Creative writing demands everything the model can do and more, and most problems are still unsolved. It&#x27;s <i>extremely</i> heavy reasoning-wise, more so than coding (check e.g. Engram paper for some insights), but also needs good scattered retrieval, careful subjective training for prose quality, character, and human likeness, a ton of facts baked in, and much much more. Mode collapse is not solved. No LLM does creative writing well but historically only the absolute largest models were able to do write anything complex more or less convincingly and were creative enough.
        • dragonwriter8 hours ago
          &gt; I am no writer but to me it seems like writing is one of the easiest tasks with barely any logic or reasoning<p>Virtually all logic or reasoning is, in one way or another, part of the support for writing. It’s what separates actual writing from generating nonsense that happens to fit grammar rules.<p>The specific details depend on the domain, of course, but I can’t see how anyone familiar with the output of writing can think that there is little logic or reasoning in doing it well.
          • Azantys40 minutes ago
            Of course there is logic but its nowhere near the complexity of math or programming
        • fc417fc80212 hours ago
          &gt; with barely any logic or reasoning<p>I take it you enjoy works of literature with inconsistent world building?<p>Or do you mean professional as opposed to creative writing? Because the bar is even higher for that.
          • Azantys38 minutes ago
            The wording wasnt very good I ment compared to programming or math the amount of logic and reasoning is small (Research level math hardly compares to writing a book in raw reasoning and logic). And I thing the smaller models have enough &quot;intelligence&quot; to write coherent with logical world building, but only the big models can truly do hard math and programming work
        • BoorishBears13 hours ago
          The largest model I&#x27;ve post-trained in the last 2 years of working on this problem was Kimi 2.5 at 1T parameters.<p>The simplest way I&#x27;d put it is, teaching a model to write coherently (follow rules, patterns, etc.) is easy enough: just use teacher forcing. Teaching a model to write creatively is easy enough: just use RL and punish it for not being creative.<p>Teaching a model to write well and creatively takes learning two partially opposing objectives that spike the learning requirements in ways that smaller models really struggle with.
          • fc417fc80212 hours ago
            &gt; is easy enough: just use RL and punish it for not being creative.<p>How are you scoring creativity in an unsupervised manner? That seems anything but easy.
            • BoorishBears12 hours ago
              Did you try reading the whole comment?<p>Once creativity is being measured in isolation, getting multiple responses from the model is enough to measure creativity a ton of different ways: wordfreq to identify overused phrases, getting multiple responses for the same prompt and promoting the least similar as preferred for policy optimization, etc.<p>But that&#x27;s of limited use for stuff like getting diverse names and such. You want creativity and coherency, and if you just punish the model for using an overused phrase, the first thing it does is strongly learn a new overused phrase (or gibberish).<p>(Also I don&#x27;t think you mean unsupervised. You probably mean without humans [since LLMs struggle to judge creativity], but that&#x27;s not what unsupervised means.)
              • fc417fc80211 hours ago
                I did read the the full comment and I did in fact mean exactly what I wrote when I used the term &quot;unsupervised&quot;. I think the condescension does nothing but get in the way. Try extending the benefit of the doubt.<p>&gt; enough to measure creativity a ton of different ways ...<p>The things you listed seem more like temperature than creativity to me. At this point it occurs to me that this is likely yet another case of highly misleading technical jargon. Suffice to say that truly creative writing requires something entirely different than unusual sentence structure - in fact it doesn&#x27;t require unusual phrasing at all.<p>Re unsupervised, it seems the misunderstanding here follows naturally from the previous difference in word meaning. Hopefully you see the difficulty of scoring long form answers for the creativity of the underlying ideas, as well as the impossibility of using a labeled dataset to train on such a criteria.
                • BoorishBears10 hours ago
                  The first thing I read from you was a sardonic browbeating in response to the exact comment I gave an earnest response.<p>And even in domains that lean heavily on &quot;usual phrasing&quot;, like technical writing, human writing has notably higher perplexity compared to another LLM&#x27;s outputs: <a href="https:&#x2F;&#x2F;www.sciencedirect.com&#x2F;science&#x2F;article&#x2F;abs&#x2F;pii&#x2F;S1076633225000170" rel="nofollow">https:&#x2F;&#x2F;www.sciencedirect.com&#x2F;science&#x2F;article&#x2F;abs&#x2F;pii&#x2F;S10766...</a><p>With such a low baseline for what&#x27;s unusual, you <i>do</i> need to get the LLM writing unusual phrases relative to its baseline. Otherwise you get things like repeated n-grams and overused constructs (&quot;it&#x27;s not X it&#x27;s Y&quot;), and suddenly the output is predictably not perceived as creative by humans even if you were to insert some otherwise creative or novel premise.<p>Getting the model to break out of that baseline without disrupting the model&#x27;s ability to follow technical rules, maintain logic and reasoning, etc. is the difficult part.<p>-<p>Also you&#x27;re again saying unsupervised then following up with descriptions that sure sound like you&#x27;re referring to RL and supervised learning respectively this time. (supervised learning can improve creativity by the way
                  • fc417fc8025 hours ago
                    &gt; Getting the model to break out of that baseline without disrupting the model&#x27;s ability to follow technical rules, maintain logic and reasoning, etc. is the difficult part.<p>Sure, that is also somewhat challenging and is necessary to get human sounding prose. However doing so is not sufficient to produce &quot;creative&quot; literature by any reasonable metric.<p>&gt; you&#x27;re again saying unsupervised then following up with descriptions that sure sound like you&#x27;re referring to RL and supervised learning respectively this time.<p>Are you sure it isn&#x27;t you who is confused about the usage of those terms? I merely suggested that both preparing and making use of labeled data (ie supervised learning) seemed like it would prove quite difficult here. Quoting from wikipedia (<a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Unsupervised_learning" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Unsupervised_learning</a>):<p>&gt; Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data.
  • ponyous24 minutes ago
    Interesting way to show off a model last on every benchmark. Not sure any other lab is doing this
  • andai16 hours ago
    Novita is offering free Hy3 on OpenRouter until July 21st<p><a href="https:&#x2F;&#x2F;openrouter.ai&#x2F;tencent&#x2F;hy3:free" rel="nofollow">https:&#x2F;&#x2F;openrouter.ai&#x2F;tencent&#x2F;hy3:free</a><p><a href="https:&#x2F;&#x2F;x.com&#x2F;novita_labs&#x2F;status&#x2F;2074158304159510819" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;novita_labs&#x2F;status&#x2F;2074158304159510819</a>
  • mattstir9 hours ago
    This is admittedly an aside from the content of the post itself, but... why do so many mobile sites insist on preventing zooming in and seem to share the same incredibly buggy image zooming? It&#x27;s quite frustrating.
  • Catloafdev18 hours ago
    Curious how people feel about this compared to DS4 Flash, given they are pretty close in size. Also curious how well it holds up to heavy quantization.<p>DS4 Flash can currently run reasonably well on systems with ~96gb+ RAM, I wonder if Hy3 can compete there.
    • tarruda17 hours ago
      &gt; given they are pretty close in size<p>One thing that might not be obvious about about DSV4 is how much innovation the Deepseek team implemented in its architecture. When llama.cpp fully supports its lightning indexer, the full 1M context will only require about 6G of RAM. So even though they are similar in size, I believe Deepseek will be much more efficient in that regard.<p>&gt; I wonder if Hy3 can compete there<p>Highly depends on how well Hy3 is resilient to quantization. DSV4 is useful even at 2-bit quants.
      • throwa3562622 hours ago
        I have been telling people exactly this the last few months.<p>We have not seen the full power of deepseek v4 yet.
    • UncleOxidant18 hours ago
      That&#x27;s a 2-bit quant of DS4 flash. You&#x27;re probably better off running Qwen3.6-27B at Q8.
      • girvo13 hours ago
        Having heavily evaluated both antirez’s ds4 flash and Qwen 3.6 27B at FP8 and Q8: it depends. The quantised Flash is better in a number of tasks despite running much slower on my DGX Spark-alike.<p>27B is amazing for its size but has some surprising limits when used for longer agentic coding sessions, especially if you’re using tools that are outside the stock standard web tech stuff: it really isn’t good at Relay, for example.
      • spmurrayzzz18 hours ago
        I think its good advice to test both on your own evals for sure, but the MoE parameters are already natively FP4 in ds4. Dropping to 2bpw isn&#x27;t as big of a loss as it seems (and as corroborated by antirez&#x27;s work).<p>Its also only 13B active, so your decode speed would be nearly 2x that of Qwen3.6-27B. So there are other latent benefits as well.
        • verdverm17 hours ago
          z-lab has been dropping dflash addons for a lot of models<p><a href="https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;z-lab&#x2F;dflash" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;z-lab&#x2F;dflash</a><p>I&#x27;m running the qwen3.6-27B + dflash on a spark and tgen is way up, but keep the draft count low, acceptance rate is terrible beyond half a dozen and it requires more memory
          • bjconlan11 hours ago
            Wow thanks for pointing this out! This is actually what I was hoping would happen when the deepspec stuff dropped! And having zlab create these confirms my bias that I think open models are the way.
      • Catloafdev18 hours ago
        For most coding or agentic tasks, Qwen 3.6 27B likely outperforms, yes.<p>For &#x27;general intelligence&#x27;, DS4 Flash seems to be a noticeable step up still.
      • sosodev18 hours ago
        I suspect it would depend on the task. DS4-flash does, as previously mentioned, handle quantization very well. Even at 2-bit it&#x27;s still very coherent.
      • FuckButtons17 hours ago
        qwen 27b at q8 is slower and worse than ds4 at q2 in my experience.
      • ckocagil17 hours ago
        Isn&#x27;t Q8 way overkill these days? I see many graphs showing Q4 or Q5 having less than %1 deviation. Nvidia&#x27;s NVFP4 Qwen quantization should be even better due to its better training methods.
        • walrus0112 hours ago
          Q8 isn&#x27;t overkill if you have sufficient RAM to fit the whole model, and you care about quality. There&#x27;s a number of people who have enough hardware to fit exactly one 27B to 35B size Q8 model and not more than that, so if you can fit the whole thing in Q8, no reason to use Q4 or Q6.
        • kgeist9 hours ago
          Qwen3.6 below Q8 often can&#x27;t exit a reasoning loop (until it hits max output token count), forgets to insert a tool call, often mistakenly inserts them inside the thinking block... It&#x27;s still usable though.
        • spmurrayzzz16 hours ago
          When orgs&#x2F;bencmarks claim 1% deviation, in most cases that means measuring perplexity loss on datasets like wikitext or c4. Even if the loss is calculated via KLD or similar, its not a good proxy for whats actually degradaing at the task level across an entire rollout.<p>And for MoEs, very small amounts of loss can mean you&#x27;re flipped to entirely different experts (this is also a problem more broadly with numerical stability issues too).
        • mdgld17 hours ago
          It depends on model size I think, but yeah, from my understanding at ~30B and below Q6 or even Q4 will get you 95%+ of the way there
        • nullc9 hours ago
          Careful with those graphs, they&#x27;re usually evaluating the model on KLD on relatively short transcripts. When you&#x27;re running with 100k token contexts and the model running close loop a difference that looks small in terms of KLD may be quite substantial.<p>I&#x27;m not aware of any great benchmarks that work by giving it a live agentic harness and a number of realistic tasks that take most of the context window to accomplish and evaluate success rate and tokens to completion... but that&#x27;s what you&#x27;d really want to use to judge different quantization levels.
      • reinitctxoffset11 hours ago
        [dead]
    • nmfisher10 hours ago
      I was playing with Hy3 via openrouter yesterday (and I&#x27;ve also been using DS4 Flash&#x2F;Pro as a daily driver since I cancelled my Anthropic sub a week ago).<p>I&#x27;ve found DS4 Flash to be very temperental (via Claude Code). The speed is great, but it often builds a completely wrong mental model and charges off down the wrong path. I find myself needing to rein it in regularly (and also compact the history, which undercuts the whole cache price advantage).<p>Hy3 isn&#x27;t as fast, but so far it seems to stay on track much more reliably than DS4 Flash. It also doesn&#x27;t seem to degrade as much with longer context. I&#x27;m not sure what the real pricing is, but I feel like it&#x27;s a very competitive model.<p>As an aside, I also nabbed a 50m token pack for LongCat 2.0 to give it a whirl. Not free, but it&#x27;s so cheap they&#x27;re basically giving it away. Very impressed too - seems roughly on par with Hy3. Not frontier-level intelligence, but a dependable workhorse that can navigate a codebase well and can reliably execute what you tell it to do.
    • wolttam18 hours ago
      Hy3 lacks the DSv4 architecture&#x27;s KV Cache efficiency.<p>Whereas I can run DSv4 Flash on a pair of DGX Sparks and have enough memory left over for 3M tokens of KV cache, with Hy3 (quantized to FP4), there is only room for ~130K tokens of KV cache.
      • ignoramous18 hours ago
        Lower context window notwithstanding, Hy3&#x27;s coding benchmarks hold their own against <i>DeepSeek v4 Pro</i> &amp; <i>MiMo v2.5 Pro</i>. That&#x27;s quite something for a model priced like <i>DeepSeek v4 Flash</i> &amp; <i>MiMo v2.5</i> (for non-cached tokens), which are 3x cheaper than their respective <i>Pro</i> variants.
        • wolttam18 hours ago
          It&#x27;s impressive indeed. I would also expect the next checkpoint of DSv4 Flash to come in somewhere at this level (DeepSeek has had over 2 months to continue training since it released).<p>It&#x27;s exciting that the open models continue to get better and more efficient across the board!
    • nunodonato18 hours ago
      DS4-Flash is not only &quot;significantly&quot; smaller, it will also benefit from a lot more speed thanks to DSpark
      • Catloafdev18 hours ago
        299B for Hy3 vs 284B* for Flash<p>Edit: fixed, got bad info
        • nshotton18 hours ago
          flash is 284b isnt it? <a href="https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;deepseek-v4-flash" rel="nofollow">https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;models&#x2F;deepseek-v4-flash</a>
          • Catloafdev18 hours ago
            Oh, it is. I was looking at the Huggingface repo which listed the lower number at the top of the page, looks like that&#x27;s wrong.
            • nunodonato17 minutes ago
              Same, I was relying on that. Weird!
    • Computer013 hours ago
      I don’t like DS4 in my experiences with it I still prefer qwen locally and glm on api
  • nshotton18 hours ago
    This model is shockingly small for how capable it is. its a little bit bigger than deepseekV4 flash but around as capable if not more on some benchmarks than V4 pro, i wouldnt be surprised if this becomes a popular local model.
    • andai18 hours ago
      I&#x27;ve been wondering about that. GLM-5.2 is also half the size of DeepSeek V4 Pro. (But costs roughly twice as much.)<p>I looked into DeepSeek&#x27;s architecture a little bit and the main focus was how can we save as much money as possible. They did a lot of cost cutting with the attention mechanisms. This allowed them to offer an insanely cheap price even on massive contexts, but seems to have come at the cost of performance?<p>At least, that&#x27;s my guess, when I see smaller models costing more and outperforming, I think, &quot;they must have denser attention?&quot;
      • wgd18 hours ago
        The current Deepseek V4 Pro is still just their initial preview AFAIK, with the &quot;real&quot; model release rumored to come later this month. GLM-5.2 might be outperforming simply because it&#x27;s had more post-training on top of the GLM-5 base.
        • VulgarExigency13 hours ago
          If the &quot;final&quot; release of Deepseek V4 Pro outperforms GLM-5.2 while maintaining the current Deepseek price, then it&#x27;s going to be a marvel.
    • nunodonato18 hours ago
      hardly, its still quite big unless by &quot;local&quot; you mean people that spend many thousands on rigs :)
      • nshotton18 hours ago
        Yeah i shouldve been more clear, a model of this size could run on 2 dgx sparks so out of the range of a lot of the typical consumer sure, but I think there is definitely a market for that size
    • IshKebab18 hours ago
      &gt; Hy3 has 295B parameters in total. To serve it on 8 GPUs, we recommend using H20-3e or other GPUs with larger memory capacity.<p>I would.
  • thot_experiment18 hours ago
    I feel like I&#x27;m taking crazy pills with hy3, it&#x27;s either benchmaxxed to hell and back or skill issue on my part but I&#x27;d rather use dense gemma. I don&#x27;t think there&#x27;s a single model that&#x27;s wasted more of my time in recent memory.
    • nik7361 hour ago
      Gemma 4 31B is underrated. It surprises me a lot.
    • SwellJoe17 hours ago
      The Hy3 preview has been a mediocre performer in my benchmarks of security auditing with models, and yes, it is outperformed by Gemma 4 (31b soundly beats it, the MoE does slightly better, even at 4-bit quantization when using the QAT version). Qwen 3.6 27b also beats it.<p>I&#x27;ll try it again now that it&#x27;s out of preview and has been updated with more post-training. It presumably can&#x27;t be worse, so maybe it&#x27;s better enough to compete with a 31b model.
      • andai16 hours ago
        For context Hy is 295B, roughly on par with DeepSeek V4 Flash (284B).<p>I haven&#x27;t tested it yet so I cannot comment on the quality (nor the comparison with 10x smaller (!) models)
        • thot_experiment15 hours ago
          There are extreme diminishing returns in real world performance as models get bigger. 10x bigger might mean 5-10% better on benchmarks, a margin that can easily mean it&#x27;s functionally equivalent in real world use or even a worse performer depending on the context it&#x27;s being used in, and how good you are at providing meaningful context.<p>Of course the bigger model embeds more knowledge, but when neither model has the knowledge necessary to perform the task, hy3 makes idiotic decisions all the time whereas gemma 31b has a decent hit rate.<p>hy3 feels like someone who&#x27;s read a lot of books and says the right words but has nothing of substance between their ears, gemma feels like a reasonably intelligent person who doesn&#x27;t understand the domain, the latter is muuuch easier to work with than the former.
          • SwellJoe15 hours ago
            Gemma 4 is the first really small model that feels smart, to me. I mean, Qwen 3.6 is arguably better at some coding tasks. But, Gemma 4 has shockingly good reasoning for a small model. Even the tiny 12B, at 7GB on disk in the 4-bit QAT quant, feels like a really big model of a couple years ago. It&#x27;s a good tool user, can search the web (when given the appropriate skill or MCP), has good vision capabilities, and pretty good prose.<p>I&#x27;ve only used the Hy3 preview, so I don&#x27;t want to judge too harshly, yet. But, I wasn&#x27;t very impressed with it a couple of months ago.
  • docheinestages17 hours ago
    What we really need is a breakthrough in inference or LLM architecture to allow running GLM-5.2-level models at the size of Qwen 3.6 27b or smaller on consumer devices like a 48GB Macbook Pro, and at least at 100 tokens&#x2F;second. My hypothesis is that a smaller, less capable but faster model paired with a good harness can run for longer and brute force its way out to solve problems that the bigger models can one-shot.
    • redwood17 hours ago
      That would be great during the winter months
      • docheinestages17 hours ago
        Why not! And free hot showers during the summer. Just looking at the progress made since 2023, I don&#x27;t think the LLM architecture we have today is the most efficient. We need creative game developers to start making LLMs.
        • fc417fc80212 hours ago
          Unfortunately I don&#x27;t expect indie game devs can afford high spec hardware at current prices.
    • cyanydeez17 hours ago
      im more expecting the harness to be a literal LLM, Like how you put vibration dampeners on all kinds of mechanical structures
  • minraws18 hours ago
    I tried out the model it&#x27;s pretty great, better than ~~gpt5.4~~ gpt-5.4-mini perhaps, atleast close enough to sonnet 5 in performance that I didn&#x27;t notice much of a gap.<p>Not really at gpt 5.5 tier though, and probably below glm 5.2...<p>But most of all it just works for me for most things I tried and it&#x27;s exceedingly cheap so there is no reason not to use it, if you need a foss model.<p>Edited: gpt-5.4-mini not the base gpt-5.4
    • cbg018 hours ago
      Hy3 DeepSWE - 28%<p>GPT5.4 xhigh DeepSWE - 52%<p>A lot of contaminated benchmarks in the blog post about Hy3, needs real testing though I have a distinct feeling it&#x27;s benchmaxxed like a lot of Chinese models.
    • theplumber18 hours ago
      I think you’ve got the models wrong…gpt-5.4? I doubt there is any open source mode matching it. Maybe in a year
      • minraws18 hours ago
        Yeah I meant gpt-5.4-mini, but GLM 5.2 is pretty close to gpt-5.4 base, and much better than it when it comes to design stuff.
      • mgrandl18 hours ago
        GLM 5.2 already matches GPT-5.4 easily.
  • baalimago1 hour ago
    Worse than GLM-5.2, much more expensive than deepseek v4. Tough bargain.
  • rw24 hours ago
    There are now 3 tiers of competition:<p>-Fable + Gpt 5.6 Sol<p>-Opus + Gpt 5.6 Terra + Grok 4.5 + Muse Spark 1.1<p>-Open Chinese models: GLM + et family<p>The economics is on the Fable tier people are willing to spend a lot on it and on the Open tier you have to give it away to drive usage. The bottom tiers are also getting more and more competitive.
  • toephu217 hours ago
    This site can’t be reached <a href="https:&#x2F;&#x2F;hy.tencent.com&#x2F;research&#x2F;hy3" rel="nofollow">https:&#x2F;&#x2F;hy.tencent.com&#x2F;research&#x2F;hy3</a> is unreachable.
  • james2doyle18 hours ago
    Been using this and GLM 5.2 back and forth. I like the speed of Hy3. Also seems very happy to follow instructions. Still haven’t found any open models that follow instructions as good as Mimo v2 pro though
    • SwellJoe17 hours ago
      MiMo v2.5 Pro is very spiky, in my experience. Sometimes excellent, sometimes mediocre. Weirdly high non-deterministic behavior. Run the same task three times, get three different results. I mean, they&#x27;re all rolling dice for the next word, but MiMo seems to run hot on the randomness dimension in my benchmarks.<p>But, it performs very well for its size. I just looked it up, and it&#x27;s much smaller than I thought it was when I was testing it. 310B A15B is tiny for how well it performs. I guess that explains why it&#x27;s so cheap.
      • SwellJoe13 hours ago
        Er, actually, Pro is a big 1T model (which makes more sense given how well it does, sometimes). Regular MiMo is small, and I haven&#x27;t tested it.
  • throwaway202718 hours ago
    Quite interesting to see them and Meta and others release before OpenAI supposedly is to release GPT 5.6 today, would it be better to release it before or after? Calm before the storm type of thing?
  • antfarm16 hours ago
    I was expecting a new release of the Hy language (<a href="https:&#x2F;&#x2F;hylang.org" rel="nofollow">https:&#x2F;&#x2F;hylang.org</a>).
    • wormius15 hours ago
      Yes, I was gonna say... <i>Today, I am this old.</i>
  • handzhiev18 hours ago
    It&#x27;s a very good model for this size and price. I tried it with a couple of small tasks - just an year ago this would be the level of the leading models.
  • 301275616 hours ago
    Congrats, the trial chat is QR locked. The AI companies are spoiled and get crazier every day (not only this one, US companies as well).
  • doawoo18 hours ago
    That UI demo page is… really quite janky.
  • assimpleaspossi16 hours ago
    I would never use any product that can&#x27;t explain on its own front page what it is and why I should use it.
    • andai16 hours ago
      It&#x27;s like a computer, but you can talk to it.
      • assimpleaspossi13 hours ago
        If they can&#x27;t tell me what it is or how to use it then they can&#x27;t explain to me how to install it or anything else properly so forget it. I&#x27;m not wasting my time trying to figure it out.
  • nunodonato18 hours ago
    Very impressive model for its size
  • jszymborski15 hours ago
    I&#x27;m sorry but what on earth is going on with that bar chart, the bars are not consistent. E.g., in the frontierscience-olympiad chart Hy3 preview scores the same as DeepSeek (70.0) but Hy3 preview&#x27;s bar is visibly lower.
  • Roguelazer16 hours ago
    Got really excited for a minute that the long-standing [Hy](<a href="https:&#x2F;&#x2F;hylang.org" rel="nofollow">https:&#x2F;&#x2F;hylang.org</a>) project had had a release, but it&#x27;s just some confusingly-named LLM. Shame.
    • linzhangrun1 hour ago
      The strange names are mainly initials from Chinese pinyin. The first generation of Tencent Hunyuan was released in 23Q3, so it is already quite a veteran.