Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Qwen 3.8

twitter.com

Post reply on HN
Print view View original

171–180 of 793 posts

Re: Qwen 3.8

#171

Post by seanmcdirmid »

tormeh wrote: post #109
ernsheong wrote: post #106

So are locally-runnable models frozen at Qwen 3.6 now :/

Is qwen 3.6 27b the best model you can run locally at the moment? Not that I have the VRAM for it, but just curious.

I actually have long discussions with Gemini about this and have wound up download a bunch of different models for different things. There is no best, just fast but worse, slow but better, agentic or not, reasoning or not great at large contexts, better world knowledge, uncensored, etc…. It’s a bit daunting actually since there isn’t really a one size fits all model that you can just use for everything.

Re: Qwen 3.8

#173

Post by Alifatisk »

I remember when they released Qwen 3.7 Plus and Max. These models behaved way different from all prior models, it became too verbose. It wrote multiple paragraphs just to answer my prompt instead of the usual concise and direct way responding to me. I didn't like that at all, and I know Gemini also had this behaviour with with the Flash series until I managed to reduce it a bit with personal instructions (in the settings on Gemini website).

I haven't tried Qwen 3.8 Max yet, looking forward to it. My hope is that its way less verbose. Another thing I experience with the Qwen models is that I do not trust their benchmark scores at all. Have anyone played with Qwen 3.8 Max and can share their experience? Which model it come close to? Sonnet 5? GLm-5? DS V4 Pro? Flash? Gemini 3.5 Flash?

Re: Qwen 3.8

#174

Post by moffkalast »

antiloper wrote: post #144

Does anyone have the privacy policy of their token plan available? Want to check if they retain/train on inputs/outputs.

Lol, lmao even.

Of course they train on literally everything they get their hands on, like everyone else. If you need privacy, that's what local models are for.

Re: Qwen 3.8

#175

Post by mattmanser »

chewz wrote: post #122
nullbio wrote: post #119

Earlier quoted context omitted.

If by Opus you mean Opus 4 and not Opus 4.8, then sure.

> If by Opus you mean Opus 4 and not Opus 4.8, then sure I meant Opus 4.8 which is rather dumb and ineffective in coding harness, especially with higher thinking levels.

Comments like this boggle my mind.

The model which everyone else raves about and is wildly successful with legions of programmers virtually demanding access while abandoning ChatGPT and Copilot in droves, is rather dumb?

Have you considered that it's more likely that you're doing something wrong?

Re: Qwen 3.8

#176

Post by cakbeslik »

Using QWEN models since 2.5. I never used the chat properly but as an API I can say they're quite good, especially when you compare with OpenAI models. Cheaper and almost same level. I will try this now also.

Re: Qwen 3.8

#177

Post by SwellJoe »

storus wrote: post #58
SwellJoe wrote: post #34

Qwen is the most censored of the Chinese models in my testing, which makes me wonder in what other ways it is compromised. Open weights doesn't really reveal what's in there. And, in my tests, existing Qwen models are not at the pareto frontier of any metric; DeepSeek V4 Pro is better, faster, and much cheaper than Qwen 3.7 Max. (DeepSeek is also among the least censored of the Chinese models.) I guess we'll see if t…

DeepSeek V4 hallucinates like crazy and often forgets explicitly mentioned parts of the context. I guess compressing tokens and cherry-picking attention comes at a cost.

I have seen occasional weird behavior that I guess could be attributed to hallucinations, but for security auditing, DeepSeek v4 Pro is among the best models I've tested, competitive with Opus 4.8 and GPT 5.5 (MiMo and GLM also did well, Qwen 3.7 Max was below all of those, though only barely), and at an order of magnitude lower cost per task.

Re: Qwen 3.8

#178

Post by weiliddat »

baq wrote: post #78
gardnr wrote: post #6

Earlier quoted context omitted.

It's hard to say what their motivation is. The Chinese firms seem to be working hard to commoditize intelligence which may be the most effective way to debase American frontier labs. And yeah: it also happens to be really good for humanity.

There’s a Twitter thread making rounds by Dean Ball about deceleration in AI development caused by open models and I can’t understand how people don’t see that it’s true: open models dismantle the frontier lab capex spend potential by reducing the training budget to zero in the limit. Tokens from different providers are not fungible, but customers are nevertheless very price sensitive and close enough is good enough,…

I read his followup tweet, and your comment, and I'm not fully convinced that open models are decelerationist. Happy to hear other thoughts on this.

Open weight AI is decelerationist from the perspective that all capital should be allocated to a market leaders for training, and that the market leader is fully invested in continuously making the models smarter, cheaper, faster for its users, or that distillation from this market leader is the main way to make progress.

We might reach a local optimum/equilibrium faster without open weight models, with leaders capturing more of the market faster to a point where further R&D isn't required due to lack of competition. I also doubt that distillation is the only/main way that open weight models were advancing AI research. We can name a few examples from DeepSeek around reasoning, context optimization, etc. I'm also unconvinced that the overall market capex on AI is lower given more competition (probably less specifically for US market capex, which is decelerationist from only the US perspective).

Re: Qwen 3.8

#179

Post by Narciss »

chewz wrote: post #122
nullbio wrote: post #119

Earlier quoted context omitted.

If by Opus you mean Opus 4 and not Opus 4.8, then sure.

> If by Opus you mean Opus 4 and not Opus 4.8, then sure I meant Opus 4.8 which is rather dumb and ineffective in coding harness, especially with higher thinking levels.

I can’t believe that anyone would actually think this. This

Re: Qwen 3.8

#180

Post by cmrdporcupine »

ernsheong wrote: post #158
tormeh wrote: post #109

Earlier quoted context omitted.

Is qwen 3.6 27b the best model you can run locally at the moment? Not that I have the VRAM for it, but just curious.

Yes it's between this and Gemma 4 31B which is much slower, but looks like it won't ever get an upgrade. I have to conclude that the MoE variants are unreliable, and MTP sometimes just can't get tricky formatting right.

For whatever reason prefill (on my DGX Spark) is faster with the Gemma models than Qwen 3.6 models of similar size. On vLLM anyways. Likely just deeply tuned code contributed to vLLM by Google?

vLLM gives me ~7000+ tok/sec with Gemma 4's MoE model. Vs ~6000 tok/sec for Qwen 3.6 MoE.

Post reply on HN