Sampling

#1
by d1sl1ke - opened

Big Thanks for the model!

Which sampling parameters do you recommend?

I am currently using the following

llama-server \
    -m ~/Projects/AI/models/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf \
    -ngl 20 -c 262144 -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0 \
    --temp 0.6 \
    --reasoning off

Really not sure if those are the right settings though, because I have just not had good experiences when Qwen 3.6 in general, seems too dumb to use Three.js correctly to make a Minecraft clone, at least it's not getting in loops as much. Maybe I'm not jamming it with enough MCPs and it's a Opencode issue, who knows, would like advise here Gemma has generally given me better results but everyone seems to disagree hopefully I'm just doing something wrong

@Jackrong just in case if you missed our question ☺️

Honestly getting way better results with just base Qwen, not where I expect though, I will give credit where credit is do much more eager to use tools than base Qwen though

Sorry for the very late reply.

You can generally use the official recommended sampling settings for Qwen3.6 with this model, as I did not make many changes to the sampling behavior.

In my experience, temp=0.6–1.0 and top_p=0.95 are both good choices. LM Studio usually defaults to around temp=1.0 and top_p=0.95, which gives the model more diversity in its responses.

This comment has been hidden (marked as Off-Topic)

Sign up or log in to comment