Wow, interesting. What KV cache quantization do you use?
854 karma · joined August 5, 2024
Interestingly, it also seems to tend toward self-correcting, which makes lower quantizations borderline usable. There'll be more faffing around, but still converging toward a solution. I wonder if that's a deliberate product of its RL.
It's both funny and a bit terrifying and I still can't quite believe it. It runs decently on a gaming PC! Opus 4.6 came out only 6 months ago and was then broadly considered the new SOTA by a comfortable margin! How in hell did they package capability in the ballpark of a Feb 2026 frontier SOTA into 27B?!
More importantly, what's the point of building monster-scale data centers on unprecedented amounts of debt when a more than good enough model runs on a GPU from a couple years ago?
The coming months are going to be exciting, that's for sure...
"The models at this size continue to get better at an impressive rate. We don’t need to spend half a million dollars on datacenter-class hardware just to run a competent model."
Worth noting as well that the weights come with an MTP layer that seems particularly accurate, to the point it can give you up to 6-8 correctly predicted tokens consistently enough to be useful. Which obviously boosts its speed enormously.
I find it difficult to believe how good this model is. It feels like it's lagging heavyweight frontier models by less than a year, and it runs on your PC.
(Which I still think is bonkers, FWIW.)
Opus 4.6 is an especially interesting comparison point, I think, because it was a step change; IMO it's when LLMs became serviceable for coding. Yeah, pre-4.6 models did output code, and that code often superficially worked; and bringing it up to production standards still generally meant rewriting it entirely. Opus 4.6 is when that changed.
From my early tests, it's looking like the public benchmarks are not misleading, and Qwen 3.8 somehow got there too, by and large. I've got a few personal tests. One is a mid-complexity one-shot, purposefully underspecified. Beyond a few minor bugs that it could easily fix once pointed out, Qwen 3.8 largely aced it. There are a bunch of things I'd improve, but that was true of Opus 4.6's output too, and by and large the code is clean and well structured. Also worth noting that I'm running Qwen 3.8 fairly aggressively quantized to fit in VRAM; I'd expect tighter results still from the full weights.
Another test I ran is a variant of a common puzzle with an additional structural constraint that makes the usual solution inapplicable, so the model has to actively turn away from the well-known solution and construct a new one that takes the constraint into account. I've never seen a home model pass that test. Kimi K3 passes it, GLM 5.2 passes it (painstakingly). Qwen 3.8 struggles a lot... but does arrive at a correct solution. First time I see a home model do so.
I haven't yet tested it on long multi-turn scenarios. In my experience, that's where pocket models are weakest against heavyweight ones, especially when quantized.
That aside, it does seem like Qwen 3.8 can, in fact, trade blows with Opus 4.6. I don't know yet if it could replace it, and my money would be on no, but I may well be wrong about that considering how weirdly capable it is.
Interestingly, Qwen 3.8's MTP layer is uncannily accurate too. It still gave me good results up to 6 to 8 predicted tokens, which boosts its speed so much it's competitive with Qwen 3.6 MoE. So that's another bizarrely impressive thing about it.
And given all of the above, I do think that the trillions of dollars invested into OpenAI and Anthropic are becoming harder and harder to justify.
And meanwhile, punishing companies that knowingly cause harm for profit is broadly beneficial and should be a no-brainer. It also doesn't have anything to do with hypocritical age verification laws, which companies like Facebook are actively pushing to shift the responsibility away from themselves. I'd rather we supported the beneficial approach over the anonymity-destroying one, if it's all the same to you.
Because, per the linked research, in present-time USA that weight is ~zero.
Which is especially interesting considering the USA's narratives about itself. And how, it seems to me, its population is so invested in said narratives that it will defend those even in the face of contrary research. You'd think a culture so invested in its self-determination would be angry about this result and endeavor to do something about it.
You might argue that Americans are just less capable of supporting themselves but I reject that argument firmly. The situation really reads like systemic causes. And I think cherry picking numbers to convince ourselves otherwise does more harm than good.