I would have liked to see a bit more on the theory side of things, explaining optimal weight and inference splits, actual issues with existing drivers, etc instead of what’s essentially just a recipe.
I can understand the joy of running things yourself, and can also see the privacy aspect. However, I pay ~3$ per 1M/tokens for that model on Openrouter, and it's not even quantized. A refurbished 3090 and a 5080 will set you back well over 2k, not to mention the electricity to run them...
I just bought a $25 chinese 2x Oculink card and two Minis Forum DEG1, had some spare PSUs lying around, and just installed two cards on each.
It works.
I saw that there is also a 4x Oculink card, but i don't know it that will work, too.
That's almost exactly my setup and I'm very happy with its performance.
I noticed recently that I started to prefer my local Qwen3.6 35B A3B and pi agent over Claude Code.
Both fail at different tasks, and Qwen more so than Claude.
But the way Qwen fails is much more straightforward. In writing tasks Qwens hallucinations and bullshitting are much easier to spot because it doesn't have the sleek vocabulary and wordsmithing skills to disguise its ignorance.
In coding tasks that Qwen can't solve it often just goes into a tool calling doom loop that the pi harness can catch, whereas Claude attempts ever more convoluted and creative things just making more and more mess that takes forever to clean up.
I think part of the story is that the tasks for which I use AI are fairly simple and maybe don't need a frontier model. But I wonder if "proper" developers had similar experience?
I have said this before as well: these top-of-the-line models write clever, convoluted code. The code looks intelligent from above, but is a maintenance headache. Makes entire thing fragile for future developments on top of it.
The smaller models, especially the aforementioned ones, they fail much more, but, do not write that insanity of the code. They do simple, non-clever coding like humans do. Much easier to maintain and build upon.
Qwen-3.6-27b is a wonderful model. Exceptionally good for it's size, and excellent in general as well. And with mtp available now, it can run at 60+ tps on a single 3090... this is roughly 30% faster tgs than most of the hosted ones being served from giant data-centers.
Not having a lot of experience with this, I ask a naive question: is there a world where you can take your local LLM and hook it up to Claude and get more Claude-like results from your local model? Obviously, there are going to be material differences in how these perform, but are we getting close to a place where this is viable? I imagine that the answers are a combination of “not yet” and “yes but it’s a lot slower” and “yes but there is actually little point to doing this because ‘what Claude gets you’ is highly baked into anthropic’s models and that’s part of what you’re paying for.”
I know the big labs like to pretend that their models are trillion parameter. But how likely is that really to be the case when Qwen 3.6 35B A3B gets so close to their performance? Seems that with the best research applied, best training data, they'd be able to top the charts with a 60B model quite easily.
Frontier models are still better (everyone would use them if it was cheap). Open source models are capable on even non "simple" problems but I trust them less, even though I usually write plans for all changes, and they are worse at debugging. I recently converted my homelab to nixos and let's just say Deepseek failed and Fable did great (the night before getting killed)
>In writing tasks Qwens hallucinations and bullshitting are much easier to spot because it doesn't have the sleek vocabulary and wordsmithing skills to disguise its ignorance.
Can't wait until we just remove the language from the LLMs for accuracy and efficiency
80tp/s with 5080 3090 combo is wild. I’ve been working with a 4090 and two Tenstorrent p150 cards, and manage only about 30 tps utilizing all three for qwen3.6 27b q8. Guess I got more optimization to do.
Would like to see the perf of their setup with and without mtp and ngram speculative decoding though, as well as parallel decode performance (once llamacpp mtp plays well with multiple slots).
Being in California electricity alone puts this non-competitive with just paying a cloud though.
I get 28tps for Qwen3.6 27B on a Ryzen AI Max 395+, with enough spare memory to run another two small models on the side. 60tps for 35B. Am surprised this is not more common.
Do you get anything useful out of your 4090 (I have one too)? Local cloud sounds like a fun idea but I just don’t see how it competes against OpenAI/Anthopic
On Apple Silicon, with MLX-LM, I am getting 20 tok/s with Macbook Max M5.
Not sure how it compares to llama.cpp performance.
In any case, while it is noticeably slower than this Nvidia RTX setup, being able to run such models on laptop is wild. Though, it heats my laptop rapidly.
It does come with one tiny little issue: it now draws 700W on full load. Just a single 5080 is enough to measurably heat up a room when loaded (320W draw at the wall on mine), and with that amount of power flowing through, you better have a good PSU as well as checking your power plugs themselves, these are going to get HOT when your entire setup is basically drawing 1kW.
I bought two 3080/20gb and one of those MACHINIST X99 mainboards as well (one with two full x16 pcie slots) those boards come with a xeon cpu included (for the pcie lane support) it set me back 800 euros total (had a spare psu, ssd and mem in a drawer) and now im also happily running 80tk/s Qwen 3.6 Q8 (MTP).
Would you mind giving these a try and let me know how they work for you? I’d imagine you would get better results and the latter will fit on a single GPU.
The recommended values for Qwen 3.6 in thinking mode is `--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.00`, and `--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00` for coding/tool calling tasks, and for non-thinking, `--temp 0.7 -top-p 0.8 --top-k 20 --presence-penalty 1.5 --min-p 0.00`.
The options listed are none of these.
Also, the recommended Qwen MTP settings are `--spec-type draft-mtp --spec-draft-n-max 2`. 3 is not good on Nvidia hardware under different workloads. You can also add `ngram-mod`, but after `draft-mtp`; however, default `ngram-mod` settings aren't well tuned, and you want `--spec-ngram-mod-n-min 12 --spec-ngram-mod-n-max 16 --spec-ngram-mod-n-match 6` (defaults are 48, 64, 24; the ratio is good, the magnitude is suboptimal).
Sits in silence, watching China as they innovated a new type of ultra-thin gpu board and calling it 5090 "Turbos." Still waiting for Shenzhen listings to post a 5090 official verified with VBIOS crack...
I tried implementing qwen through openrouter and deepinfra. Even without thinking, I had to wait 60s+ for the full result, where haiku or flash would be done in 5 or 6 seconds.
It is absolutely mind blowing to see some of the responses here. Open source, run-your-own, pay for nothing, we’re-all-nerds-that-buy-the-hardware-anyways ethos seems basically dead.
I guess I’m getting old. I own two 16gb cards and I use them for models, for gpu-pasthru for gaming, 3d model rendering, etc. 14 year old me is mortified at this community.
With qwen3.6-35b-a3b-mtp using lm-studio on RTX 3090, I was getting 120tokens/s. The mtp (multi token prediction) is the key.
I tired coding with Pi and it was much faster than Claude, but for any not-straightforward tasks, it did so so. Either looping itself or not realising easy to spot constraints.
But for exploring codebases and asking questions about big stuff I find it better due to sheer speed.
34 comments
[ 2.6 ms ] story [ 117 ms ] threadI noticed recently that I started to prefer my local Qwen3.6 35B A3B and pi agent over Claude Code.
Both fail at different tasks, and Qwen more so than Claude.
But the way Qwen fails is much more straightforward. In writing tasks Qwens hallucinations and bullshitting are much easier to spot because it doesn't have the sleek vocabulary and wordsmithing skills to disguise its ignorance.
In coding tasks that Qwen can't solve it often just goes into a tool calling doom loop that the pi harness can catch, whereas Claude attempts ever more convoluted and creative things just making more and more mess that takes forever to clean up.
I think part of the story is that the tasks for which I use AI are fairly simple and maybe don't need a frontier model. But I wonder if "proper" developers had similar experience?
The smaller models, especially the aforementioned ones, they fail much more, but, do not write that insanity of the code. They do simple, non-clever coding like humans do. Much easier to maintain and build upon.
Qwen-3.6-27b is a wonderful model. Exceptionally good for it's size, and excellent in general as well. And with mtp available now, it can run at 60+ tps on a single 3090... this is roughly 30% faster tgs than most of the hosted ones being served from giant data-centers.
Do you get the speed of the 5080 with the memory of the 3090?
Can't wait until we just remove the language from the LLMs for accuracy and efficiency
Would like to see the perf of their setup with and without mtp and ngram speculative decoding though, as well as parallel decode performance (once llamacpp mtp plays well with multiple slots).
Being in California electricity alone puts this non-competitive with just paying a cloud though.
NVIDIA GeForce RTX 5080: https://flopper.io/gpu/nvidia-geforce-rtx-5080-16gb
NVIDIA GeForce RTX 3090: https://flopper.io/gpu/nvidia-geforce-rtx-3090-24gb
On Apple Silicon, with MLX-LM, I am getting 20 tok/s with Macbook Max M5. Not sure how it compares to llama.cpp performance.
In any case, while it is noticeably slower than this Nvidia RTX setup, being able to run such models on laptop is wild. Though, it heats my laptop rapidly.
If you're not power limiting in nvidia-smi, start.
https://huggingface.co/easiest-ai-shawn/Qwen3.6-27B-ExCal-EX...
https://huggingface.co/easiest-ai-shawn/Qwen3.6-27B-ExCal-Mi...
Do be sure to use dflash and/or mtp for the draft:
https://huggingface.co/turboderp/Qwen3.6-27B-MTP-exl3
https://huggingface.co/turboderp/Qwen3.6-27B-DFlash-exl3
The options listed are none of these.
Also, the recommended Qwen MTP settings are `--spec-type draft-mtp --spec-draft-n-max 2`. 3 is not good on Nvidia hardware under different workloads. You can also add `ngram-mod`, but after `draft-mtp`; however, default `ngram-mod` settings aren't well tuned, and you want `--spec-ngram-mod-n-min 12 --spec-ngram-mod-n-max 16 --spec-ngram-mod-n-match 6` (defaults are 48, 64, 24; the ratio is good, the magnitude is suboptimal).
Of abliterated Qwen 3.6 27B models, huihui's ends up being the worst. Try heretic instead. https://huggingface.co/mradermacher/Qwen3.6-27B-uncensored-h...
I guess I’m getting old. I own two 16gb cards and I use them for models, for gpu-pasthru for gaming, 3d model rendering, etc. 14 year old me is mortified at this community.
90 t/s for 27B Q8 256k context
260 t/s for 35B-A3B Q8 256k context
I tired coding with Pi and it was much faster than Claude, but for any not-straightforward tasks, it did so so. Either looping itself or not realising easy to spot constraints.
But for exploring codebases and asking questions about big stuff I find it better due to sheer speed.
Though if you're buying a X570 board I'd do crosshair viii dark hero - no buzzy chipset fan and can also do 2x8