OpenClaw stumbles on local Qwen3.5

This looks less like a bad model and more like a brittle agent integration stack. Reasoning distills, Ollama imports, and OpenClaw’s tool-calling path all have to line up, and a tiny mismatch can make a fast GPU feel broken.

• –Ollama can import GGUFs directly with `FROM ./model.gguf` and `ollama create`; the Modelfile is the blueprint, not a separate artifact you need to hunt down online.
• –Jackrong’s model card is built around structured `<think>...</think>` reasoning, so template or stop-token mismatches can derail output quickly.
• –OpenClaw only surfaces models that report tool support during auto-discovery, so a custom Qwen distill can look "missing" or brittle even when Ollama itself is fine.
• –OpenClaw’s guide also recommends tool-capable models like Qwen 2.5 or Llama 3.3, which suggests a reasoning-distilled Qwen3.5 variant may not be the most forgiving agent backend.
• –The "compute spikes instead of staying solid" symptom reads like bursty retries or context thrash, not steady token generation.