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

Qwen3.6-35B-A3B: Agentic coding power, now open to all

qwen.ai

Post reply on HN
Print view View original

531–540 of 563 posts

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#531

Post by altruios »

onlyrealcuzzo wrote: post #530
altruios wrote: post #525

I have moved through the local models at this size. This one is by far the most capable. I've tried various versions of gemma4.26b, various versions of qwen3.5-27/35b (qwopus's galor),nemotron,phi,glm4.7. This one is noticeably better as an agent. It's really good at breaking down tasks into small actionable steps, and - where there is ambiguity - asks for clarification. It's reasoning seems more solid than gemma4, t…

How does it compare to CC Opus Max?

I try not to use publicly hosted models, and I avoid the SOTA data harvesting machine... so I can not compare. I can compare only to local models. And this one feels like a decently significant leap compared to 3.5 or gemma4.

I see there is now a distilled reasoning model version on hugging-face. I may look into that, but I have not seen a need to reach for that change yet either...

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#532

Post by tmaly »

mncharity wrote: post #347
tmaly wrote: post #275

What is the min VRAM this can run on given it is MOE?

Fwiw, with its predecessor's Qwen3.5-35B-A3B-Q6_K.gguf, on a laptop's 6 GB VRAM and 32 GB RAM, with default llama.cpp settings, I get 20 t/s generation.

That is pretty solid, I have a 2070 with 8GB VRAM and 64GB RAM, but I haven't run too much. I regret not getting a 3090 back when I built this machine.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#533

Post by bigyabai »

wren6991 wrote: post #449
bigyabai wrote: post #186

Earlier quoted context omitted.

taps the sign Unified Memory Is A Marketing Gimmeck. Industrial-Scale Inference Servers Do Not Use It.

On M5 Pro/Max the memory is actually just attached straight to the GPU die. CPU accesses memory through the die-to-die bridge. I don't see the difference between that and a pure GPU from a memory connectivity point of view. Wrt inference servers: sure, it's not cost-effective to have such a huge CPU die and a bunch of media accelerators on the GPU die if you just care about raw compute for inference and training. App…

If you're going to do unified memory, that's the way to do it, in addition to using higher-bandwidth RAM and padding out your GPGPU hardware. Nvidia realized this almost a decade ago, and Apple is being dragged through the mud to learn the exact same $4 trillion lesson.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#534

Post by Tamitami2000 »

kylehotchkiss wrote: post #252
zozbot234 wrote: post #40

Earlier quoted context omitted.

This is just one model in the Qwen 3.6 series. They will most likely release the other small sizes (not much sense in keeping them proprietary) and perhaps their 122A10B size also, but the flagship 397A17B size seems to have been excluded.

How many people/hackernews can run a 397b param model at home? Probably like 20-30.

According to this blog (https://kaitchup.substack.com/p/lessons-from-gguf-evaluation...) the UD_IQ2_M quants are quite strong (rel. error to the base is very low), so it's around 120GB of RAM needed, while the experts can be loaded into VRAM and the rest offloaded into system RAM. It's a high end consumer PC, sure, but not unaffordable. For example, I got an older rig with a RTX 6000 ADA (48GB VRAM), 128 GB RAM and a Threadripper, which runs this quant offloaded at 20 tps

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#535

Post by Tamitami2000 »

jubilanti wrote: post #283
kylehotchkiss wrote: post #252

Earlier quoted context omitted.

How many people/hackernews can run a 397b param model at home? Probably like 20-30.

You can rent a cloud H200 with 140GB VRAM in a server with 256GB system ram for $3-4/hr.

Can you tell me where? I used runpod before, but they don't have systems like that.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#536

Post by culi »

bmitc wrote: post #465
culi wrote: post #326

Earlier quoted context omitted.

the more I look at these images the more convinced I become that world models are the major missing piece and that these really are ultimately just stochastic sentence machines. Maybe Chomsky was right

> that these really are ultimately just stochastic sentence machines I thought that's exactly what they are?

No, they have "attention". There is unique logic going on in the deep layers of the neural network.

Even the standard introductory exercise artificial neural networks, handwritten digit recognition, already shows deeper understanding. These simple networks take in raw pixels and somewhere in the many layers recognize "curves" and "edges" and then "circles" and "boxes" and whatnot and eventually "digits".

I think there's a genuine debate about whether or not this is a form of intelligence. I think the oversimplified argument of them just being stochastic sentence machines mostly comes from people who don't understand how they work. But I also think there's a much more nuanced version of this argument offered by people like Chomsky that should be taken seriously

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#537

Post by culi »

mastermage wrote: post #467
bmitc wrote: post #465

Earlier quoted context omitted.

> that these really are ultimately just stochastic sentence machines I thought that's exactly what they are?

I am so perplexed what exactly where people thinking they were. Its nothing else than highly sofisticated statistics.

Do you know of any other statistical model that can "hallucinate". They clearly have emergent capabilities that come from scale that are absent in any other statistical model we've ever dreamt up.

We know that LLMs build complex internal representations of language, logic, and concepts rather than just shallow word-counting.

If you deny that then you probably have an elementary understanding of how they work. Not even Chomsky denies that. The real argument imo is whether those internal representations constitute an actual "understanding" of the world or just flatten out to something much less interesting.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#538

Post by Tamitami2000 »

aliljet wrote: post #73

I'm broadly curious how people are using these local models. Literally, how are they attaching harnesses to this and finding more value than just renting tokens from Anthropic of OpenAI?

I'm using forge code (https://forgecode.dev/) with various local and cloud models and I really like it. MiniMax 2.7 is really great with it, and the new Qwen 3.6 35B A3B feels much stronger, after some testing, than the 3.5 version. Check some harness benchmarks. Forge outperforms Claude Code with Opus by a big margin.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#539

Post by kanemcgrath »

mettamage wrote: post #502
kanemcgrath wrote: post #356

I have been using Qwen3.5-35B-A3B a lot in local testing, and it is by far the most capable model that could fit on my machine. I think quantization technology has really upped its game around these models, and there were two quants that blew me away Mudler APEX-I-Quality. then later I tried Byteshape Q3_K_S-3.40bpw Both made claims that seemed too good to be true, but I couldn't find any traces of lobotomization doi…

who do you compare it against qwen3.5 27b?

I haven't ran 27b that much because it only runs at like 2 tokens/sec on my computer.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#540

Post by jononor »

Zopieux wrote: post #399
vlapec wrote: post #362

Earlier quoted context omitted.

No need to hope; it is inevitable.

Is it inevitable though? Open-weight models large enough to come close to an API model are insanely expensive to run for con/prosumers. I'd put the “expensive” bar at ≥24GB since that's already well into 4 digits, which gives you quite many months of a subscription, not including the power will for >400W continuous. Color me pessimistic, but this feels like a pipe dream.

A decent amount of software developers and gamers do spend 3000 USD on a PC. That kind of hardware is going go get more and more capable over time wrt genAI models.

Of course there will always be a gap to frontier closed hosted models. It is not an either or proposition.

Post reply on HN