alt Hacker NewsI feel like these types of benchmarks mostly test what the mainstream tells you is possible and or works. You rarely get to hear the different implementations of inference engines, different types of setups (e.g. hybrid ones or older hardware) and their pros and cons. Also it seems these were tested with very minimal token usage.
Currently on my 9 year old laptop with a GTX 1070 Max-Q 8GB vram, 16 gb ddr4 ram and i7 8750h. I am running a finetuned Qwen 3.5-35B-A3B (Tiel-Coder-35B-A3B-GGUF-MTP) with 4-bit quant on a llamacpp fork (Turboquant) with 131k context at around 16-19 tk/s and 350-500 tk/s prefil using Pi harness. It is amazing for writing scripts and doing medium coding tasks and summarizing text for me. After around 100k context it goes down to 9 tk/s but that is to be expected.
If you ever ask any frontier LLM or these types of benchmarks, they would probably say that the best model I can run on my system is a crappy 9B model or something, which is just completely false. So I really wonder how much more you can pull out from more powerful systems and much better open weight models.
Replies
The A3B part of that model means that each "expert" in the mixture of experts (MOE) model use 3B parameters. At 4-bit quant, you can fit 2 parameters per byte. That gives you 1.5GB per expert (plus a bit for overhead), which makes it usable on an 8GB card.
It's all about trade-offs: how much of the model you want to load into VRAM; how much VRAM you want to reserve for context; how much quantization you/the model can tolerate; how many tokens/second you can tolerate; whether you want to use the VRAM for anything else (graphics, Text-to-Speech, etc.); etc.
Thanks for sharing! I would love for some go to source for this type of insider info from the people who actually have tried. (The best I've found so far are some guys on Twitter/X like Alex Fateev (https://x.com/superalesha) and AJ (https://x.com/ItsmeAjayKV))
I'm personally sitting on an extremely limited Geforce GTX 1660 6GB, and would like to see how far one could go even with such a crappy card.
At least it was quite amazing to see the difference between CPU-based and even this crappy GPU-based inference when trying out some tiny models in jan.ai. It's still like orders of magnitude faster at least.
This tool has very tiny quants available for each model, and defaults to the tiny quants, so I'm not so sure you... tried it?
It sounds great and I hope local LLM/AI will be more powerful and available even on old hardware.
I tried llama.cpp with offloading "all what could be" to my GPU (GTX 750 Ti 2 GB). I could run local Qwen2.5-Coder quite fast (I do not remember but something like 25-30 t/s). However this model generates bad code (I tried Python and JavaScript). So, for my setup it is better to chat with free ChatGPT and do copying-pasting between web chat and code editor.
Is it even worth for trying something on my GPU? What do you think?
And yes, I use Claude Code (Opus 5.5) and Codex in my job but for personal stuff I would prefer trying free solutions (it is the reason I tried local LLM). I do not want to buy subs, at leats for now.
I need to play around with this.
I am running on a Ryzen 5950X, 64GB DDR4 Ram and a Radeon RX 6700 XT with 12GB... I don't expect wonders, but I have between 11 and 120 (!) t/s.
I told it an open ended question: "What would you improve in this file at {location} in {bigger project}" and the context window seems to be too small. It's running in circles.
I must be doing something wrong.
Again, I am not expecting wonders and I am totally willing to accept that "I am holding it wrong".
llama-server \ -m ~/models/tiel-coder/Tiel-Coder-35B-A3B-MTP-UD-IQ4_XS.gguf \ -ngl 20 \ -c 16384 \ --jinja \ --alias tiel-coder
using it in opencode
{ "$schema": "https://opencode.ai/config.json", "provider": { "llamacpp": { "npm": "@ai-sdk/openai-compatible", "name": "Local llama.cpp", "options": { "baseURL": "http://127.0.0.1:8080/v1" }, "models": { "tiel-coder": { "name": "Tiel-Coder 35B A3B", "limit": { "context": 16384, "output": 4096 } } } } }, "model": "llamacpp/tiel-coder" }
Locally the MOE based models get a lot more out of the hardware than a dense model. Qwen 3.8-27B is the current golden child if you have >24GB VRAM but because its a dense model it is pretty slow to run and requires a lot of hardware. A lot of the guides are very much focussed on getting the dense models to run but in practice the MOE models with few active parameters are just faster and easier to run without too much loss of intelligence, they behave a lot more like their peak size than the active parameter size.
RAM is going to be the big bottleneck for MOE (at least for now) as it will determine the quantisations you could run.
On a 3080 10GB and 32GB of RAM I can run the qwen-next coder model and that is a 127B model and it gets 40-50 Token/s with 65k context. That is a lot more usable than a Quantised Qwen3.8-27B in my experience both because its faster and more capable.
I am looking forward to Qwen4 soon, I am hoping we get some more 27-35B MOE models, those are going to be really good for local hardware.