As this is a dense model and it's pretty sizable, 4-bit quantization can be nearly lossless. With that, you can run this on a 3090/4090/5090. You can probably even go FP8 with 5090 (though there will be tradeoffs). Probably ~70 tok/s on a 5090 and roughly half that on a 4090/3090. With speculative decoding, you can get even faster (2-3x I'd say). Pretty amazing what you can get locally.
> As this is a dense model and it's pretty sizable, 4-bit quantization can be nearly lossless
The 4-bit quants are far from lossless. The effects show up more on longer context problems.
> You can probably even go FP8 with 5090 (though there will be tradeoffs)
You cannot run these models at 8-bit on a 32GB card because you need space for context. Typically it would be Q5 on a 32GB card to fit context lengths needed for anything other than short answers.
Even bumping up to 16-bit K cache should fit comfortably by dropping down to 64K context, which is still a pretty decent amount. I would try both. I'm not sure how tolerant Qwen3.5 series is of dropping K cache to 8 bits.
> You cannot run these models at 8-bit on a 32GB card because you need space for context
You probably can actually. Not saying that it would be ideal but it can fit entirely in VRAM (if you make sure to quantize the attention layers). KV cache quantization and not loading the vision tower would help quite a bit. Not ideal for long context, but it should be very much possible.
I addressed the lossless claim in another reply but I guess it really depends on what the model is used for. For my usecases, it's nearly lossless I'd say.
Turboquant on 4bit helps a lot as well for keeping context in vram, but int4 is definitely not lossless. But it all depends for some people this is sufficient
4-bit quantization is almost never lossless especially for agentic work, it's the lowest end of what's reasonable. It's advocated as preferable to a model with fewer parameters that's been quantized with more precision.
Yeah, figure the 'nearly lossless' claim is the most controversial thing. But in my defense, ~97% recovery in benchmarks is what I consider 'nearly lossless'. When quantized with calibration data for a specialized domain, the difference in my internal benchmark is pretty much indistinguishable. But for agentic work, 4-bit quants can indeed fall a bit short in long-context usecase, especially if you quantize the attention layers.
This isn't the first open-weight LLM to be released. People tend to get a feel for this stuff over time.
Let me give you some more baseless speculation: Based on the quality of the 3.5 27B and the 3.6 35B models, this model is going to absolutely crush it.
Not at all, I actually run ~30B dense models for production and have tested out 5090/3090 for that. There are gotchas of course, but the speed/quality claims should be roughly there.
> Btw as an aside, we didn’t announce on Friday because we respected the IMO Board's original request that all AI labs share their results only after the official results had been verified by independent experts & the students had rightly received the acclamation they deserved
> We've now been given permission to share our results and are pleased to have been part of the inaugural cohort to have our model results officially graded and certified by IMO coordinators and experts, receiving the first official gold-level performance grading for an AI system!
I think this is them not being confident enough before the event, so they don't wanna be shown a worse result than competitors. By being private they can obviously not publish anything if it didn't work out.
This reminds me of when OpenAI made a splash (ages ago now) by beating the world's best Dota 2 teams using a RL model.
...Except they had to substantially bend the rules of the game (limiting the hero pool, completely changing/omitting certain mechanics) to pull this off. So they ended up beating some human Dota pros at a psuedo-Dota custom game, which was still impressive, but a very much watered-down result beneath the marketing hype.
It does seem like Money+Attention outweigh Science+Transparency at OpenAI, and this has always been the case.
Limiting the hero pool was fair I'd say. If you can prove RL works on one hero, it's fairly certain it would work on other heroes. All of them at once? Maybe run into problems. But anyway you'd need orders of magnitude more compute so I'd say that was fair game.
It's not even close to the same game as Dota. Limiting the hero (and item) pool so drastically locks off many strategies and counters. It's a bit hard to explain if you haven't played, but full Dota has many more tools and much more creativity than the reduced version on display. The behavior does not evidently "scale up", in the same way that the current SotA of AI art and writing won't evidently replace top-level humans.
I'd never say it's impossible, but the job wasn't finished yet.
That's akin to saying it's okay to remove Knights, or castling, or en passant from chess because they have a complicated movement mechanic that the AI can't handle as well.
Hero drafting and strategy is a major aspect of competitive Dota 2.
When your goal is to control as much of the world's money as possible, preferably all of it, then everyone is your enemy, including high school students.
I am still surprised many people trust him. The board's (justified) decision to fire him was so awfully executed that it lead to him having even more slack
Maybe not a popular sentiment here on HN but I cancelled my Kagi subscription (9+ months) just recently. Increasingly, most of my queries/search have been through LLMs and Google search is just fine (and even better for restaurants, places, and the like). I don't think the improved search experience is worth the subscription anymore.
It’s not a euphemism - every outage, including the 99.9% that don’t end up on HN gets a postmortem document written about it, which is almost always a fascinating discussion of the technical, cultural and organisational situation that led to an unexpected bad thing happening.
Even a few years ago senior management knew to stay the fuck out except for asking for more info.
I think it's most illustrative to see the sample battles (H2H) that LMArena released [1]. The outputs of Meta's model is too verbose and too 'yappy' IMO. And looking at the verdicts, it's no wonder by people are discounting LMArena rankings.
> This will mark the first experimental model with higher rate limits + billing. Excited for this to land and for folks to really put the model through the paces!
Traditionally at Google experimental models are 100% free to use on https://aistudio.google.com (this is also where you can see the pricing) with a quite generous rate limit.
This time, the Googler says: “good news! you will be charged for experimental models, though for now it’s still free”
Right but the tweet I was responding to says: "This will mark the first experimental model with higher rate limits + billing. Excited for this to land and for folks to really put the model through the paces!"
I assumed that meant there was a paid version with a higher rate limit coming out today
> The bottleneck then becomes how to self-host the finetuned model in a way that's cost-effective and scalable
It's not actually that expensive and hard. For narrow usecases, you can produce 4-bit quantized fine-tunes that perform as well as the full model. Hosting the 4-bit quantized version can be done on relatively low cost. You can use A40 or RTX 3090 on Runpod for ~$300/month.
reply