3 ms·
Tried yesterday on my own laptop (a UltraCore 7 255H without dedicated GPU,with 32 GB RAM), it wasn't even starting thinking, even on a small context window (65
by madduci 3mo ago
Tried yesterday on my own laptop (a UltraCore 7 255H without dedicated GPU,with 32 GB RAM), it wasn't even starting thinking, even on a small context window (65k)
- aphroz 3mo agoI think not much can run without a dedicated GPU
- noduerme 3mo agoWhat's the story with Mac laptops? Worth a try?
- rawland 3mo agoYes. mtplx runs it at 25 tok/sec on a M4 Max with 48GB RAM.
- selcuka 3mo agoThe author tested in on an M5 laptop too: > It feels pretty slow on both the M5 Mac and the DGX Spark.
- cyberrock 3mo agoDense ones like this are more bandwidth-hungry, so you want to try MoE ones like Qwen3.6-35B-A3B (35 Billion params but only 3 Billion Active) or Gemma 4. Unfortunately it seems like we might not be getting a 3.8 MoE.
- madduci 3mo agoTill now I was using successfully Qwen 3.5 and Gemma 4 at a reasonable speed
- mobelkh 3mo agowere you running the MoE models? those perform better speed wise
- pyrale 3mo agoThere is no way you would run a dense 27b model on that spec. I ran 3.6 27b on a 64gb ram, 24 gb vram, and it felt like the lower limit for this model with a decent context window. If you want a better experience, maybe wait for either a moe model (like 3.6 35b A3) or a model with less parameters (like 9b). Qwen has been releasing those in the past, so maybe we’ll have them for 3.8 too.
- DanielHB 3mo agoFrom my experience if it doesn't fit on vram it is rarely worth to bother except for a few narrow tasks. For example make an essay about something where you don't actively engage with the LLM after the initial prompt. So mostly one-shot prompts.
- madduci 2mo agoThe issue was caused by Ollama. I've tried again with llama.cpp and mounting the igpu device correctly, I get 18-23 tk/s on the Iris Xe card under Debian, MUCH better than before.
- kzrdude 3mo agoI think (maybe I missed something) that identical size and quant versions of Qwen 3.5 and 3.8 should run at the same speed. It’s the exact same architecture.
- madduci 3mo agoTried the 4 Bit versions. It loads, bit the <thinking> output isn't even coming out.
- pdyc 3mo agoi have same 255h and i was able to run it with low token speed 6-8tg/s with approx similar context window 60k
- madduci 3mo agoInteresting. What are you using? I was using ollama
- NamlchakKhandro 3mo agoDon't bother with plans
- pdyc 3mo agoi am using llama.cpp with pi harness on linux(cachyos). some of the things like coopmat are disabled in linux due to intel driver issues otherwise performance would be even better.
- mdp2021 3mo agoHave you tried with different amounts for the "reasoning_effort (xhigh|medium|low)" parameter? Or the "<|think_xhigh|> | <|think_low|> | <|think_off|>" tags: apart from this template detail, it is not immediately clear if reasoning_effort is deterministic (API) or is prompt engineering.
- madduci 3mo agoNo, good point. I will have to tried it
- petu 3mo agoWhat was your prompt length? It's possible it was just processing it and it's likely not fast on your setup.
- madduci 3mo agoReally small (<100 tokens), I wanted to test its capabilities
- tylerKorhonen 3mo ago> it wasn't even starting thinking Probably stuck in prompt processing which is compute bound especially for iGPUs. You've mentioned 3.5 - but it's actually the same model the only differences are training and implicit MTP support (affects prompt processing - can be disabled)