News for hackers. Stuff that matters.
Re:Qwen4-Max by pu_pe on Monday November 16, 2026 @03:30AM (#50515340)
I think Alibaba is not aiming for the frontier on this class. Their previous Qwen3-Max was released in September, right around when Opus 5.5 was coming out, and it was nowhere close to it as well. However, one or two tiers below (Flash, 27B) they really shine and are SOTA or close to it.

DeepSeek has a similar problem, I feel like their full model is hardly even worth using, while the Flash tier is absolutely amazing.

Re:Qwen4-Max by mrklol on Monday November 16, 2026 @03:38AM (#50515392)
I am unsure. Why shouldn’t they? Why even work on a big model (especially training time) if “nobody will use it as competitors are better”? We’re still talking about nearly 2.5T params here. Deepseek’s pro model is still heavily smaller and like you said isn’t directly worth using, obviously it has the open source / self hosting advantage.

Imo there’s a need and they try to push here, it’s also giving them advantages for their smaller models but if the big models never perform they will have a bad time or maybe should even stop.

Re:Qwen4-Max by cyanydeez on Monday November 16, 2026 @03:53AM (#50515480)
My use of qwen3.8-flash-next suggests they're trying to make a low cost model that knows how to program and they think the "total addressable market" is serving more people at lowest inference costs possible.

Think of it like a hind brain+fore brain discussion. Instead of focusing on making some ultra model that can do everything but costs a ton, you have your smaller models gather the bits and feed it better data

Agentic coding is the target then. If they want their work force to not care about token gen, that's what you do.

As an analog, they're making a drone swarm while US is making 5th Gen fighters. Completely different battle fields.

Re:Qwen4-Max by anentropic on Monday November 16, 2026 @05:45AM (#50516148)
The 'frontier' incorporates two dimensions though, usually score x cost, so it's not just the best score.

Meaning models below the frontier either cost more for same score or score worse for same cost.

Re:Qwen4-Max by cyanydeez on Monday November 16, 2026 @10:24AM (#50518566)
I think the argument is like MacOS vs windows vs linux. AI models from anthropic/openAI are windows and MacOS; they're plat forms for "skilless" to start on and of course you can gravitate towards the development end. They're bloated in that sense.

Open models follow the linux goals which is small tools that chain together. If I can get a small model to understand how to program for another model or do the programming bits for a large model, then the trade off is flexibility. You can imagine your giant AI that can do anything, like windows/Mac is to your benefit, but there's a clear flexibility line to how these models act.

Even if it takes longer for my local model to find something, that doesn't mean it's taking a long time because it's doing wrong things.

Honestly, I think a lot of people using SOTA don't understand what you can do with a proper engineering harness; we all live in a yellow submarine.

Re:Qwen4-Max by pu_pe on Monday November 16, 2026 @06:34AM (#50516426)
Yeah, I feel like models like Qwen3.8 27B are making clear trade-offs, like having very little world knowledge baked in. I believe those are intentional, so they can focus more on other areas. I guess that penalizes them on these benchmark aggregators.
Re:Qwen4-Max by namr2000 on Monday November 16, 2026 @08:08AM (#50517222)
If you're aiming to make the best possible small model, you still need to train a larger model and then do logit-level distillation from it. This is where most of the gains in small models have come from.
Re:Qwen4-Max by thibsrn on Monday November 16, 2026 @08:47AM (#50517575)
[delayed]
Re:Qwen4-Max by bluebands on Monday November 16, 2026 @01:01PM (#50520653)
> As opposed to the big American shops which focuses solely on performance, they have operational costs and other factors in their constraints and requirements

This is not true, OpenAI and Anthropic compete strongly at the lower end. See Haiku 5.5 and Luna

Re:Qwen4-Max by seanmcdirmid on Monday November 16, 2026 @11:54AM (#50519756)
> DeepSeek has a similar problem, I feel like their full model is hardly even worth using, while the Flash tier is absolutely amazing.

DeepSeek also agrees, and tried to withdraw their pro model but got pushback so they are offering it even if they are not encouraging new use cases for it.