#I'm lookig for a good LLM for my openclaw. I'm using Qwen3.5-4B
1 messages · Page 1 of 1 (latest)
Let me check the docs paths so I can give a clean answer instead of guessing.
For OpenClaw, I would not use a small local model as your main coding model on that machine.
Short version:
- Qwen 4B-class is too small for reliable coding + tool use in OpenClaw.
- With a 5070 Ti-class GPU, local is still workable, but mostly for lighter prompts and with more latency.
- If you want the best experience, use a strong cloud model as primary and keep a local model as fallback/privacy option.
- If you want local anyway, use the largest full-size model you can fit, and avoid tiny/heavily quantized builds.
OpenClaw’s docs are pretty blunt about this:
- a single 24 GB GPU only really works for lighter prompts
- local setups should use the largest model you can run
- smaller models are worse for tool calling and prompt-injection safety
Docs: