Pro Tip für alle Qwen 3.5 Agentic Coder

When Qwen released the Medium Series of Qwen 3.5 models last week, I was immediately hooked.

I replaced the GLM-4.7 Flash model with the Qwen 3.5 Flash model and was very happy.

I was only a little disappointed with the coding. I often had to make manual adjustments.

Then I found the recommendation from Unsloth AI. They recommend reducing the temperature from 1.0 to 0.6 for coding tasks. This changed everything. After that, the results from Qwen-3.5-Flash were shockingly good.

I can only recommend this model. I use it with opencode and achieve fantastic results on my MacBook.

./llama-server \
    --model Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf \
    --alias "Qwen3.5" \
    --jinja  \
    --temp 0.6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.00 \
    --repeat-penalty 1.0 \
    --ctx-size 131072 \
    --port 8001