I have been overly critical and arguing in bad faith about your writing in the past. As well as negative towards you, which in turn was breeding a bad environment.
While I dont really enjoy LLMs, you did help me realize my unreasonable feelings as well as realize the occupation (and the joys I got from it) is essentially dead from it’s previous iteration and that I should let go and just join in the “I’m doing it for the money and attention” crowd. I will still just hand code my own projects and not use LLMs when I can.
I think it’s cool you started the pelican meme however useful it really is even if only aesthetically.
As of today, it has fallen to 8/9th on the rankings. I don'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.
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.
DS4 Flash can currently run reasonably well on systems with ~96gb+ RAM, I wonder if Hy3 can compete there.
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.
> I wonder if Hy3 can compete there
Highly depends on how well Hy3 is resilient to quantization. DSV4 is useful even at 2-bit quants.
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't as big of a loss as it seems (and as corroborated by antirez's work).
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.
I'm running the qwen3.6-27B + dflash on a spark and tgen is way up, but keep the number low, acceptance rate is terrible beyond half a dozen and it requires more memory
Isn't Q8 way overkill these days? I see many graphs showing Q4 or Q5 having less than %1 deviation. Nvidia's NVFP4 Qwen quantization should be even better due to its better training methods.
Hy3 lacks the DSv4 architecture's KV Cache efficiency.
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.
Lower context window notwithstanding, Hy3's coding benchmarks hold their own against DeepSeek v4 Pro & MiMo v2.5 Pro. That's quite something for a model priced like DeepSeek v4 Flash & MiMo v2.5 (for non-cached tokens), which are 3x cheaper than their respective Pro variants.
It'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).
It's exciting that the open models continue to get better and more efficient across the board!
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/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.
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.
I've been wondering about that. GLM-5.2 is also half the size of DeepSeek V4 Pro. (But costs roughly twice as much.)
I looked into DeepSeek'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?
At least, that's my guess, when I see smaller models costing more and outperforming, I think, "they must have denser attention?"
The current Deepseek V4 Pro is still just their initial preview AFAIK, with the "real" model release rumored to come later this month. GLM-5.2 might be outperforming simply because it's had more post-training on top of the GLM-5 base.
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
I tried out the model it's pretty great, better than ~~gpt5.4~~ gpt-5.4-mini perhaps, atleast close enough to sonnet 5 in performance that I didn't notice much of a gap.
Not really at gpt 5.5 tier though, and probably below glm 5.2...
But most of all it just works for me for most things I tried and it's exceedingly cheap so there is no reason not to use it, if you need a foss model.
A lot of contaminated benchmarks in the blog post about Hy3, needs real testing though I have a distinct feeling it's benchmaxxed like a lot of Chinese models.
I feel like I'm taking crazy pills with hy3, it's either benchmaxxed to hell and back or skill issue on my part but I'd rather use dense gemma. I don't think there's a single model that's wasted more of my time in recent memory.
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?
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
It'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.
I tried the preview model 41 days ago and got a pelican with a "change pelican color" button: https://static.simonwillison.net/static/2026/hy3-preview-pel...
While I dont really enjoy LLMs, you did help me realize my unreasonable feelings as well as realize the occupation (and the joys I got from it) is essentially dead from it’s previous iteration and that I should let go and just join in the “I’m doing it for the money and attention” crowd. I will still just hand code my own projects and not use LLMs when I can.
I think it’s cool you started the pelican meme however useful it really is even if only aesthetically.
As of today, it has fallen to 8/9th on the rankings. I don'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.
https://openrouter.ai/tencent/hy3-preview
https://openrouter.ai/deepseek/deepseek-v4-flash
DS4 Flash can currently run reasonably well on systems with ~96gb+ RAM, I wonder if Hy3 can compete there.
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.
> I wonder if Hy3 can compete there
Highly depends on how well Hy3 is resilient to quantization. DSV4 is useful even at 2-bit quants.
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.
https://huggingface.co/collections/z-lab/dflash
I'm running the qwen3.6-27B + dflash on a spark and tgen is way up, but keep the number low, acceptance rate is terrible beyond half a dozen and it requires more memory
For 'general intelligence', DS4 Flash seems to be a noticeable step up still.
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.
It's exciting that the open models continue to get better and more efficient across the board!
Edit: fixed, got bad info
I looked into DeepSeek'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?
At least, that's my guess, when I see smaller models costing more and outperforming, I think, "they must have denser attention?"
I would.
Not really at gpt 5.5 tier though, and probably below glm 5.2...
But most of all it just works for me for most things I tried and it's exceedingly cheap so there is no reason not to use it, if you need a foss model.
Edited: gpt-5.4-mini not the base gpt-5.4
GPT5.4 xhigh DeepSWE - 52%
A lot of contaminated benchmarks in the blog post about Hy3, needs real testing though I have a distinct feeling it's benchmaxxed like a lot of Chinese models.