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

Qwen3-4B-Thinking-2507

huggingface.co

Post reply on HN
Print view View original

31–40 of 64 posts

Re: Qwen3-4B-Thinking-2507

#31

Post by whimsicalism »

Workaccount2 wrote: post #29
evilduck wrote: post #27

Earlier quoted context omitted.

Meta seems to have now stepped out of the running despite being the local LLM catalyst. Anthropic has done nothing. IBM's Granite and Microsoft's Phi are both very far behind. AWS doesn't even attempt to compete. Grok failed to make good on their promise. OpenAI only even entered the game yesterday so it's hard to tell if they're actually serious since they released such an overly censored model that isn't really bet…

China has a business model where you can lose money and it doesn't matter. The state's modus operandi is just fund things until the leader changes his mind about it. This is why the Chinese labs are so open, they don't ever need to make a profit, they just need to make good AI.

Sure, except all of these Chinese labs are attached to large, profitable Chinese tech & finance companies. But yeah, it's all unfair competition.

Re: Qwen3-4B-Thinking-2507

#32

Post by threeducks »

cowpig wrote: post #7
esafak wrote: post #6

Earlier quoted context omitted.

https://artificialanalysis.ai/leaderboards/models?open_weigh...

Compare these rankings to actual usage: https://openrouter.ai/rankings Claude is not cheap, why is it far and away the most popular if it's not top 10 in performance? Qwen3 235b ranks highest on these benchmarks among open models, but I have never met someone who prefers its output over Deepseek R1. It's extremely wordy and often gets caught in thought loops. My interpretation is that the models at the top of Artific…

OpenRouter rankings conflate many factors like output quality, popularity, price and legal concerns. They can not tell us whether a model is popular because it is genuinely good, or because many people have heard about it, or because it is free, or because the lawyers trust the provider.

Re: Qwen3-4B-Thinking-2507

#33

Post by svnt »

It is interesting to think about how they are achieving these scores. The evals are rated by GPT-4.1. Beyond just overfitting to benchmarks, is it possible the models are internalizing how to manipulate the ratings model/agent? Is anyone manually auditing these performance tables?

Re: Qwen3-4B-Thinking-2507

#34

Post by Imustaskforhelp »

smallerize wrote: post #19
gok wrote: post #2

So this 4B dense model gets very similar performance to the 30B MoE variant with 7.5x smaller footprint.

It gets similar performance to the old version of the 30B MoE model, but not the updated version. https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507

I still think that its still very commendable though.

I am running this beast on my dumb pc with no gpu, now we are talking!

Re: Qwen3-4B-Thinking-2507

#35

Post by meatmanek »

jampa wrote: post #11

I am reading this right, is this model way better than Gemma 3n[1]? (For only the benchmarks that are common among the models) ===== LiveCodeBench E4B IT: 13.2 Qwen: 55.2 ===== AIME25 E4B IT: 11.6 Qwen: 81.3 [1]: https://huggingface.co/google/gemma-3n-E4B

Reasoning models do a lot better at AIME than non-reasoning models, with o3 mini getting 85% and 4o-mini getting 11%. It makes some sense that this would apply to small models as well.

Re: Qwen3-4B-Thinking-2507

#36

Post by nisten »

If you want to have an opinion on it,

just install lmstudio and run the q8_0 version of it i.e. here https://huggingface.co/bartowski/Qwen_Qwen3-4B-Instruct-2507....

you can even run it on a 4gb raspberry pi Qwen_Qwen3-4B-Instruct-2507-Q4_K_L.gguf https://lmstudio.ai/

Keep in mind if you run it at the full 262144 tokens of context youll need ~65gb of ram.

Anyway if you're on mac you can search for "qwen3 4b 2507 mlx 4bit" and run the mlx version which is often faster on m chips. Crazy impressive what you get from a 2gb file in my opinion.

It's pretty good for summaries etc, can even make simple index.html sites if you're teaching students but it can't really vibecode in my opinion. However for local automation tasks like summarizing your emails, or home automation or whatever it is excellent.

It's crazy that we're at this point now.

Re: Qwen3-4B-Thinking-2507

#37

Post by evilduck »

Workaccount2 wrote: post #29
evilduck wrote: post #27

Earlier quoted context omitted.

Meta seems to have now stepped out of the running despite being the local LLM catalyst. Anthropic has done nothing. IBM's Granite and Microsoft's Phi are both very far behind. AWS doesn't even attempt to compete. Grok failed to make good on their promise. OpenAI only even entered the game yesterday so it's hard to tell if they're actually serious since they released such an overly censored model that isn't really bet…

China has a business model where you can lose money and it doesn't matter. The state's modus operandi is just fund things until the leader changes his mind about it. This is why the Chinese labs are so open, they don't ever need to make a profit, they just need to make good AI.

Sure. The US isn't morally any better on this though. VC backed companies and the "get users, figure out profitability later" strategy that have made outside countries unable to compete have been enabled by the US government tweaking interest rates and the supply value of the dollar. And the US backs all sorts of unprofitable things through grants, contracts, and bail outs.

The US government just hasn't yet found a reason to directly pay for a domestic country to release open models. But it's not like we're above that at all.

Re: Qwen3-4B-Thinking-2507

#38

Post by Workaccount2 »

whimsicalism wrote: post #31
Workaccount2 wrote: post #29

Earlier quoted context omitted.

China has a business model where you can lose money and it doesn't matter. The state's modus operandi is just fund things until the leader changes his mind about it. This is why the Chinese labs are so open, they don't ever need to make a profit, they just need to make good AI.

Sure, except all of these Chinese labs are attached to large, profitable Chinese tech & finance companies. But yeah, it's all unfair competition.

I don't follow

Xi Jinping has set AI as a national priority. This means that lenders (Chinese party run banks) will "lend" money to AI orgs with no financial strings attached. No business evaluation needed. This is how China does growth, they just fund it in the direction they want without much attention to profitability or returns. It's how you get billion dollar high speed rail lines that transport 50 people a day along the route.

Despite the many capitalist facets of China, it's core operations are still planned communist economy.

Re: Qwen3-4B-Thinking-2507

#39

Post by whimsicalism »

Workaccount2 wrote: post #38
whimsicalism wrote: post #31

Earlier quoted context omitted.

Sure, except all of these Chinese labs are attached to large, profitable Chinese tech & finance companies. But yeah, it's all unfair competition.

I don't follow Xi Jinping has set AI as a national priority. This means that lenders (Chinese party run banks) will "lend" money to AI orgs with no financial strings attached. No business evaluation needed. This is how China does growth, they just fund it in the direction they want without much attention to profitability or returns. It's how you get billion dollar high speed rail lines that transport 50 people a day…

That you do not follow is apparent from your comment. I don't think this will be a productive discussion - you're speaking in such broad generalities about provable facts about specific companies (Alibaba, Baidu, etc.) that it is difficult to respond. Take care.

Re: Qwen3-4B-Thinking-2507

#40

Post by Aeroi »

nisten wrote: post #36

If you want to have an opinion on it, just install lmstudio and run the q8_0 version of it i.e. here https://huggingface.co/bartowski/Qwen_Qwen3-4B-Instruct-2507... . you can even run it on a 4gb raspberry pi Qwen_Qwen3-4B-Instruct-2507-Q4_K_L.gguf https://lmstudio.ai/ Keep in mind if you run it at the full 262144 tokens of context youll need ~65gb of ram. Anyway if you're on mac you can search for "qwen3 4b 2507 mlx…

how about on apple silicon for the iphone
Post reply on HN