Hi, i haven’t seen anybody do what title above says. Idk, maybe everyone nowadays do this already :) But if not, I want to show off a little. There are my specs
12th Gen Intel® Core™ i5-12450H (12)z
GPU 1: NVIDIA GeForce RTX 3050 Mobile [Discret]
GPU 2: Intel UHD Graphics @ 1.20 GHz [Integrat]
16GB RAM DDR4
Running on cachyos (arch linux), because on Windows, proven by my tests, speed is lower (Gemma 4 E4B 40t/s on linux, and and 30t/s on windows).
I used UD-IQ4_NL quant version (13.4GB), as it seems like the best compromise between quality and size.
Using ik_llama.cpp fork due to optimizations with MoE and CPU + GPU hybrid work.
These are the flags i use
“$LLAMA_SERVER”
-m “$MODEL_PATH”
-ngl 99
-c 8000
-fa on
-ctk iq4_nl
-ctv iq4_nl
–parallel 1
-nkvo
-t 8
-tb 8
-b 256
-ub 256
-rtr
-amb 512
–no-mmap
–jinja
-mla 2
–cpu-moe
–mlock
–reasoning off
so there is very little batch size. Even 512 causes OOM. Prefill can take time when context becomes bigger.
Not all flags are actually doing something, i just tried everything i found that can help.
Doing the most - cpu-moe (offloading experts to ram), little batch size, and nkvo (offloading kv cache to ram).
Result(u can see token speed) on screenshot.
15t/s - MoE architecture saves the day!
As the result:
- The chat quality is great. Facts are solid, instruction following great too
- Model is bad on agentic tasks sadly
Great model on just medium class device with limited VRAM, and prove (at least to myself) that26B models don’t need 16GB VRAM to run PROPERLY.
The main problem now - is usable context window and prefill speed. On 8k the speed is 10t/s. Waiting for author of the ik_llama.cpp to implement turboquant to help solve the problem. Luckly he already works on that.
PS. tried running qwen3.6 35B. Again - the size is the main problem. Used Apex-i-mini version (14gb). It runs succesfully, speed is 20t/s, but quality is really bad. Will try to max out what i can on UD_IQ4_NL quantisation
UPD: UD_IQ4_NL too big, trying APEX-COMPACT
UPD 2: With a bit of tweaking here and there i balanced memory consumption on VRAM and RAM and APEX-COMPAT version of Qwen3.6 35B… attention… BLASTED with 30 tokens per second! That’s just wow. Now problem is that there is only 100mb left on RAM and i can’t even open the browser…
So for now, i connected to local server from my phone. And yeah - 30t/s. That’s crazy. But no room for context really… Need to figure something out…
It took me until now to finally dabble in these coding agents. And I didn’t realize at all how many tokens they burn through. I let it write some basic HTML & JavaScript browser game with some free OpenRouter model. I’ve done this before, just told a model to one-shot it in a single file. And now I tried OpenCode, let it ask me a few questions, come up with a plan and do an entire project structure… And it’s at one million tokens way faster than I thought. If my math is correct, that’d take my computer 2 days and nights straight at 6T/s 👀
Guess it’s really a bit (too) slow.
the problem with coding agents is simple - THERE A LOT of System promts. Promts that correct the behavior of the model in process of creating project. That is needed becase even largest models are a dumb to some degree. They forget what tools they need to use and how to use them properly. So there hidden from you system promt (i tried Cline for example - it is 11k tokens only on system prompt!) that eats context like crazy. I tried to create similar agent with tools and system promts, that save on context (my custom tool “get_overview”, instead of read_file; in mix with “search_content” tool that returns lines on search query, it can save a lot - model don’t need to read full file) and mix just a tiny beat cheetsheet to every user msg, so model don’t forget. Results were very good. Don’t know why they need spam sysprmt like that.
So i think this problem is kinda solvable on local machine