[hypfer]

After having been a happy user of Qwen3.6-27B for a few weeks, due to being away from the hardware, I'm currently forced to use Claude Sonnet 4.6

It is such a downgrade. I don't understand how that's even possible. The thing has so many strongly-held opinions I did not ever ask it for, talking just way too much and generally feeling somehow dumber.

Of course, being significantly larger, it will encode more knowledge, but that doesn't help me when I hate talking to it. And all that on top of the fact that talking with it costs real money.

I wonder what it might be that makes me hate it so much. Maybe because it doesn't see itself as a tool but almost an equal? As if its opinions would have weight.

Qwen too can act like an overeager intern, but if you tell it that it is an idiot, it will drop that ego. Not so much with Claude. In my experience, anyway.

Anyway, point is: full ack on that headline.

[ggerganov]

I haven't spent a dime on cloud inference, so cannot make a direct comparison like you. But I can 100% attest to the fact that Qwen3.6-27B is a very capable local model for coding tasks. Over the last month and a half I've been using it almost daily, either on my M2 Ultra or on my RTX 5090 box. I use it for small mundane tasks at ggml-org [0] - nothing really impressive, but definitely a helpful tool for a maintainer. I think I would be using it much more, if I didn't have to spend a lot of my time on reviewing PRs. Currently, I have a very lightweight harness - the pi agent with everything stripped (`pi -nc --offline`) and a short system prompt [1] to align it a bit with my style. About the generation speed: ~100-150 t/s on the RTX 5090 and ~40 t/s on the Mac. I definitely prefer running it on the RTX machine - it's so much faster. But for the sake of testing and getting wider experience with local configurations, I often run it on the Mac too.

[0] - https://github.com/search?q=%22Assisted-by%22+user%3Aggml-or...

[1] - https://github.com/ggml-org/llama.cpp/blob/master/.pi/gg/SYS...

[trilogic]

I also confirm that local inference is on par with proprietary cloud services (with a bit of local setup, simple agents.md and some utils skills). This local models come with tools, that's mind blowing, considering that some months ago we had to .md tools ourselves. What makes a model worth even more is "Memory". We implemented that long ago. Last time I used proprietary services was 3 months ago, don´t really need it, my subscription is going blank.

Gerganov, hope you will consider developing further the CLI cause we suffering with the server.

[kpw94]

> About the generation speed: ~100-150 t/s on the RTX 5090 and ~40 t/s on the Mac

Curious if you can share the prefill speed too?

I run locally on a crappy desktop (some AMD iGPU with Vulkan llama.cpp, 32 GB DDR4 RAM) for experimentation. I get 15 tok/s on generation for the qwen & gemma4 MoE models. I get around 150 tok/s prefill speed.

Reason I'm asking about the prefill is looking at my stats at work, I use between 20M to peaks of 300M input tokens daily. Some of those token are cached but in general, I seem to have roughly 500x more input tokens than output. So interested in prefill tok/s stats.

Huge Thank you for llama.cpp btw!!

[ggerganov]

Here are the prefill speeds:

    Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes, VRAM: 32109 MiB
  | model                          |       size |     params | backend  |  fa |            test |                  t/s |
  | ------------------------------ | ---------: | ---------: | -------- | --: | --------------: | -------------------: |
  | qwen35 27B Q4_K - Medium       |  15.92 GiB |    27.32 B | CUDA     |   1 |   pp2048 @ d512 |      3714.02 ± 10.85 |
  | qwen35 27B Q4_K - Medium       |  15.92 GiB |    27.32 B | CUDA     |   1 |  pp2048 @ d1024 |      3684.86 ± 15.21 |
  | qwen35 27B Q4_K - Medium       |  15.92 GiB |    27.32 B | CUDA     |   1 |  pp2048 @ d2048 |       3650.80 ± 8.53 |
  | qwen35 27B Q4_K - Medium       |  15.92 GiB |    27.32 B | CUDA     |   1 |  pp2048 @ d8192 |       3473.88 ± 0.97 |
  | qwen35 27B Q4_K - Medium       |  15.92 GiB |    27.32 B | CUDA     |   1 | pp2048 @ d32768 |       2754.69 ± 4.07 |

  ggml_metal_device_init: GPU name:   MTL0 (Apple M2 Ultra)
  | model                          |       size |     params | backend  | fa |            test |                  t/s |
  | ------------------------------ | ---------: | ---------: | -------- | -: | --------------: | -------------------: |
  | qwen35 27B Q8_0                |  26.62 GiB |    26.90 B | MTL      |  1 |   pp2048 @ d512 |        379.75 ± 0.21 |
  | qwen35 27B Q8_0                |  26.62 GiB |    26.90 B | MTL      |  1 |  pp2048 @ d1024 |        377.15 ± 0.35 |
  | qwen35 27B Q8_0                |  26.62 GiB |    26.90 B | MTL      |  1 |  pp2048 @ d2048 |        371.46 ± 0.91 |
  | qwen35 27B Q8_0                |  26.62 GiB |    26.90 B | MTL      |  1 |  pp2048 @ d8192 |        344.84 ± 0.41 |
  | qwen35 27B Q8_0                |  26.62 GiB |    26.90 B | MTL      |  1 | pp2048 @ d32768 |        222.42 ± 5.29 |

Btw, based on your numbers, I think our use cases are quite different. I use the agent for very targeted sessions - basically things that are clear to me how to do, just want to automate them. My workflow is usually: new session -> read this, this and this -> do that. I.e. I don't let it wander at all in the codebase, so I rarely exceed the context window.

Also, I get a lot of mileage from the ngram-based speculative decoding functionality [0] as it allows me to iterate on the implementation much faster.

[0] https://github.com/ggml-org/llama.cpp/pull/19164

[kpw94]

Thanks! Super helpful.

I do use it the same way as you're describing on personal projects at home, in a very crude manner (pasting code snippets in llama server web UI prompt. Next will attempt OpenCode)

At work I use it in similar manner with more mature tools, but the vast majority of token spend comes from a totally different workflow: "pretend the AI is a fleet of junior/intern engineer you're delegating work to", where the agent will on its own do the implementation, commit the changes etc.

It does indeed spend a lot of tokens wandering the codebase, talking to MCPs, loading skills etc.

[celrod]

What quant do you run it at? 32GB seems like cutting it close on the rtx 5090 if going 8b, but other commenters are saying 4b lobotomizes the model.

[ggerganov]

As a baseline, I run all models in Q8 [0] because I want to be confident that when I observe a problem, the root cause is not due to the quantization. However, in this specific case, I use Q8 on the mac and Q4 on the RTX machine because the latter does not fit the full context at Q8. So far, I don't have conclusive evidence that the Q4 quantization affects the quality in a significant way for this model and the tasks that I am using it for.

[0] https://huggingface.co/ggerganov/presets/blob/main/preset.in...

[fridder]

Not too shabby. I like the regular Qwen but prompt prefill on my m1max is slow as hell

[StevenWaterman]

Yep, I daily drive Qwen3.6-27B (including for work), have done pretty much since it came out. IMO it's the only (small-ish, local) model worth using, if you can run it. It might not be as good as Opus at "add X large feature" but I don't want that in a model. I want to do the thinking while it does the typing. And Qwen 3.6 27B is perfectly good at that (while in my experience models like the 35A3B and gemma are significant downgrades)

Plus, I never have to worry about rate limits, quotas, or sitting in a queue during peak time. And I can always see its full thoughts, don't have to worry about where my data is getting sent, and know it can't get secretly nerfed.

Running on 2x 3090, 500-1000tok/s prefill and 60tok/s output at Q6_K_XL with MTP on llama.cpp, 220k tokens context window (starts to get a bit dumb above 160k ish), no KV quantization

[indoordin0saur]

> And I can always see its full thoughts, don't have to worry about where my data is getting sent, and know it can't get secretly nerfed.

For this reason I wonder if local models are a potential business opportunity. Provide the service to engineering teams to give them a pre-built and setup GPU rig they can run in a closet. No need to worry about all the things you mentioned and clients can rest-assured their data isn't disappearing into a sketchy data center. There might be regulatory reasons that make on-prem setups appealing as well.

[amoshebb]

This is, as far as I know, the business model of coys like mistral and cohere

[suncemoje]

On-premise (1960-2010) -> Cloud (2010-2026) -> On-premise (2026+)?

[indoordin0saur]

I think that's overstated, but the loss of trust companies have with the big AI players is pretty serious. Not a big deal if your app is for sharing cat videos, but if you're medical or wealth management or a government contractor or the like enterprise clients really like to see good data security policies.

[suncemoje]

Agree. I also wonder how zero e.g., Claude Enterprise ZDR really is, and what their data pipeline actually looks like.

[cyanydeez]

I think the next step to anyone but overbloated USA models is to follow https://chatjimmy.ai/ with one of the qwen models. If they can mass produce something at relative cost, these would be awesome sidecars.

[Andrex]

How long have you been using it?

[iamtheworstdev]

are you running an NVLink? I have the same setup but no NVLink and it feels like it's best just splitting the 3090s to run separate models concurrently. But I also have no idea what I'm doing.

[hughw]

Just this morning I tweaked my single 3090 setup too:

  OLLAMA_FLASH_ATTENTION=1
  OLLAMA_KV_CACHE_TYPE=q8_0
  OLLAMA_CONTEXT_LENGTH=180000

and that fits in 23GB.

[edited for format]

[giancarlostoro]

> (starts to get a bit dumb above 160k ish)

If open models can ever hold roughly 600k token windows, I'll be really excited, I found that around 300 ~ 400k of Claude reading through your codebase results in better outputs. I also have Claude read official docs instead of "guessing" as to how to do something.

[StevenWaterman]

I think we'll get there. Right now it works for me, because I'm naturally pretty verbose in my prompts, and know the codebase well, so I know what it needs to look at. Plus subagents for anything exploratory.

I think deepseek v4 pro has 1m context and does pretty well up to around 600k. But if you have the hardware to run that locally, you already know

Even then if there's a smaller model with 1M context, you'll need a ton of RAM to actually run it at full 1M. I guess that's why you don't see it too much. Anyone that could run Qwen 3.6 27B with 1m context would be better off running a much bigger model with smaller context instead, in the same amount of VRAM.

In terms of optimizing further, huge context + KV quantization sounds like a terrible idea, but there's some decent innovation in sparse attention, KV cache rotation allowing Q8 to perform nearly as well as full 16-bit precision, plus some ideas around offloading KV cache to system RAM (but I'm skeptical)

[zozbot234]

DeepSeek V4 (both Flash and Pro) has very good scaling of context length wrt. RAM use, so this is not an inherent limit of LLMs in general.

[0xc133]

With yarn and rope scaling arguments for llama.cpp you could run qwen3.6-27B with 1M context… if you have enough memory to store it.

[cyanydeez]

I don't really think you're making reasonable decisions at that size; but I suppose if you're not allowed to refactor it, maybe.

I think the way these models work excludes sane behaviors the larger the context gets as each token introduces potential ambiguities between "USER" and "SYSTEM" messages leading to all the catastrophic behaviors.

Anyway, with AMD395+ I'm finding ~100k is both speed and context usefulness unless it's scoped tightly. with opencode, I manage it with dynamic context pruning: https://github.com/Opencode-DCP/opencode-dynamic-context-pru... ; then anything I touch ends up being refactored so context doesn't get bloated with unecessary functions, etc.

Obviously, this isn't compatible with certain business codebases, so I can see why bloat meets bloat.

[QuantumNoodle]

Do you have any resources on hardware necessary for running models and tweaks? I see you mention 2x 3090 and I wanted to do more search on what hardware is satisfactory for what models.

[epistasis]

> talking just way too much

OMG this is such an annoying property, just shut the hell up please, and be concise.

I suspect that this is an artifact of the thinking property, but please just summarize the thinking process far more concisely, where a single sentence answer is more than sufficient the frontier models seem devoted to going on to a minimum of 5 paragraphs and offering 3-5 new directions.

And requests to please only offer a single step at once, or single option at once, or to even stop eagerly offering future directions is really hard to prompt correctly.

And look, there I did exactly what I was complaining about...

[bityard]

I'm not sure to what degree you can influence how a model thinks, but you can definitely hide the thinking tokens and tell the model how you want it to talk to you.

For example, the Claude web UI has an Instructions field where I have told it never to congratulate or praise me for asking questions. Earlier Copilot models used a ridiculous number of emoji and bullet lists when answering literally every prompt, I told it to knock that off and prefer detailed paragraphs in prose.

Local agents/frameworks/whatever all have their equivalents for overall user preferences.

[epistasis]

Thanks for the reminder! For others looking for this setting, it is currently under User Menu (click your account name in the lower left), then "Settings", then the "General" tab there's an "Instructions for Claude" box.

Asking Claude for this provides incorrect instructions for me, so I'm guessing it moves around a lot.

[illegalsmile]

That's why you have to give claude and others directives/.md at the beginning so it doesn't go off the deep end with suggestions.

[epistasis]

Yeah, I've tried, and I'm sure somebody is going to say "skill issue" but it's not so easy to get the model to do that. Maybe it should be a SKILLS.md issue.

Edit: also, how can I stop the LLM from all this fake glazing, as if every question I have is some sort of unique genius insight, it's so damn annoying. I just got the third straight round of this while merely trying to get summarization of a PDF:

> Good question — it gets right at a real tension in the paper. Let me check the current state of actual SV-imputation efforts, since this has moved since 2020.

[bornfreddy]

I didn't try telling to be concise and stop pampering me yet (but good idea, tomorrow), however I found that instead of me writing agent instructions, it works much better if I tell claude to write instructions for itself. I do check if they make sense of course, but its wording works much better than mine.

[radium3d]

If you think about it, they're splitting the power across millions of users. Essentially, these AI companies have YOUR hardware that YOU are paying (them) for in a cabinet at some data center. This means the hardware could easily be run locally for inference for these 'big' models. It's just a problem of dynamics-- RAM is being bought in bulk by these companies through these B200 style cards, instead of sold slowly through the open public markets.

This is likely due to a combination of mass funding for the AI companies, but also they are trying to governmentally restrict which countries get access to these cards so certain countries get a head start. The only way to lock that down is to have them literally locked in their own GPU prisons (data centers). Third reason is it does make it possible to train the models faster by having them in the same data center connected directly. Having them distributed to everyone would slow down training considerably.

The current way to 'own' decent RAM and GPUs right now is through the stock market it seems.

[kitd]

Funny that coding agents have personalities, including "that colleague" you want to avoid even if you know they're probably quite good at what they do!

[derethanhausen]

I would not generalize based on experiences with Sonnet. The flagship models (Opus being the claude equivalent) are dramatically better.

[hypfer]

Opus in my experience is equally unpleasant "character"-wise, but at least it actually gets stuff done more often, so it's at least slightly more earned at that. It's still a neurotic cargo-culting dogmatic idiot, but one that at least sometimes does produce deliverables instead of only bottom-tier HN-esque opinions.

Hmm. I think I might just fundamentally disagree with Anthropic about the idea of what a "tool" should be.

[MostlyStable]

Curious if you have tried custom instructions. I was never quite as unhappy with Claude's voice as you appear to be, but there were several things I didn't like. A custom prompt fixed almost all of them.

[clickety_clack]

I think it would be very hard to convince someone to pay $100/mo to go back to Claude if they have a local model up and running, particularly now that model improvement has basically been stalled for the last 6 months. It’s so easy to set it up for yourself now too with things like LM studio. That said, there will always be unsophisticated users who can’t figure it out, so there will always be someone there to pay.

[MostlyStable]

The person I was replying to specifically said that the Claude will "encode more knowledge" and that their problem was that they didn't like talking to Claude. It sounds like they think that Claude is at least slightly more functional. And the "not liking talking to it" is probably fixable. Someone for whom a local model works, and for whom the economics make sense, should absolutely run a local model and I wouldn't try to convince them otherwise. I'm sure it's the right choice for a lot of people. But not liking the personality of Claude is probably not a great reason on its own, given the minuscule amount of effort it takes to fix.

[Scoundreller]

The third category are the occasional users that won’t have the hardware and won’t stomach a monthly fee for “unlimited” but are happy to pay-per-use.

I’d think the volume for that category would be low but LLMs aren’t just for coding.

[dghlsakjg]

I’m probably the third category. I like experimenting and trying different models and techniques. I want api access for my own apps and Claude subscriptions don’t have that.

Sure I could splash out a ton of money for a high ram Mac, but deepseek is so dirt cheap that I think depreciation on a high end machine costs more than my api spend.

Example of what I’m using it for: building a semantic database of podcast content (podcast discoverability sucks on an episode level). I need a cheap LLM, an embedder, a transcriber, none of which Claude will do.

My api costs for coding agents plus running apps are about ~$20/month, but I get more than just chat + Claude code.

If all I was doing was pumping an employers codebase through a coding agent, Claude would be the answer.

[chrisweekly]

Not everyone has the right hardware.

[clickety_clack]

I guess I’m thinking of the $100/mo users, for whom it’s probably possible to get the right hardware.

[giancarlostoro]

There's a model on Huggingface where someone takes Qwen and makes it think Opus style, and that one seems to be decent, not sure if they have the 27B variant in that style. I do wonder if you can tweak your system prompt to force Qwen to behave better?

[StevenWaterman]

You read the OP backwards, they said Sonnet is a downgrade from Qwen, and prefer Qwen's tone

[giancarlostoro]

Sure, but my argument still holds, the idea is that Qwen reasons the way that Opus on High (what is now Max or whatever?) level thinking to reason about problems instead of its standard approach.

[whythismatters]

Yes, Qwopus :) I've been pleasantly surprised by its quality

[giancarlostoro]

Seen that one too, same guy I'm thinking of too, havent had a chance to try all of their models. For anyone curious I believe the username is Jackrong on huggingface? They've got several models out on there each focused on programming from different approaches.

[zerd]

I noticed Fable was quite a bit terser, and I think it's due to changes in the system prompt [0]. They're literally saying "just give me the TLDR" and "give brief updates". You can tweak a lot of that with an AGENTS.md.

[0] https://twelvetables.blog/comparing-claude-fable-5s-system-p...

[dackdel]

what kind of hardware do you need in order to run qwen3.6-27b

[giancarlostoro]

Depends on which variant you pull down, but a single 5090 GPU (I know these are insanely expensive, but for context) could run either the Q8 or Q4_K_M version. It will not fit the 52GB version (BF16) on the other hand. So any modern Mac with a Pro or better processor and more than 52GB of RAM (don't forget VRAM for context window also matters!) would suffice, as someone else noted, probably a 128GB model would do the trick, and give you enough wiggle room to max out the context window.

My Mac only has 16GB of VRAM (20GB total - 8 is reserved for the OS) so I have to leave room for VRAM, I usually find a model that fits in 5 to 7 GB of VRAM and then max the context window as much as I can.

[iagooar]

I recommend MacBook M5 Max with 128 GB of RAM to run it comfortably and fast. If you have something like a regular M4, go with qwen3.6-35b-a3d - the Mixture of Expert architecture makes it run 2-3x faster than the 27b version.

[sbmthakur]

I could run it on 7900 XT with 64k context. You could run it more comfortably on a 24 gb vram.

[indoordin0saur]

Very curious what hardware you're running this on!

[hypfer]

The same 24GB VRAM RTX 4090 I bought to play Cyberpunk 2077 with.

Works perfectly fine in llama.cpp throwing 70+t/s at me with 128k q8 K/V context when using the IQ4_NL quant + MTP at q4 MTP K/V.

Also leaving this here because you might find it useful: https://hypfer.github.io/will-it-fit-llama-cpp/

[indoordin0saur]

Nice! Do you do anything with that compute when you're not actively using it? Is the crypto-mining hobby still worth it? I've also wondered if such expensive hardware can be rented back out to offset cost. Looks like these cards are going for as much as $4k nowadays.

[all2]

There are services where you can hook your card up and rent it out to other users. I don't know what any of them are called, but they do exist.

[dghlsakjg]

Salad.com is one. (I’m unaffiliated, just happened to come across it this week while looking for a cheap option)

[hypfer]

I've paid ~2k€ in 2023. Since I'm usually sitting next to it, I'm only using it when I want to use it. It can get quite loud and warm.

Crypto (to my knowledge at least) moved away from GPU mining. I guess you could maybe rent out GPU compute, but - being in germany - it's not worth the legal hassle. You could of course always commit tax fraud, though I wouldn't recommend that.

[esseph]

> I've also wondered if such expensive hardware can be rented back out to offset cost.

Massive legal liability. Not worth it.

[cdelsolar]

What did you call me?

[calebm]

sync/ack

[ltononro]

Well but comparing with sonnet 4.6 instead of opus 4.6,.7 or .8 doesnt make a real point I mean, pay 200 USD/month (if you have that cash, or your company has it), might not justify using local at all (unless you have some reason to suspect about data leakage)

[chrisweekly]

Why Sonnet 4.6 not Opus?