2dph/docs/reasoner.md

Reasoner bake-off (D18)

Pluggable OpenAI-compatible URL. 2dph does not ship weights. PicoClaw is not in this repo; the bake-off hits the same tool names PicoClaw would (search → get → audit from internal/httpapi.Ops).

docker compose --profile reasoner up -d reasoner
docker compose --profile reasoner exec reasoner ollama pull qwen3.5:9b
REASONER_BASE_URL=http://127.0.0.1:11435/v1 REASONER_MODEL=qwen3.5:9b \
  ./bin/reasoner/bakeoff.go --json

Host Ollama on :11434 is left alone. This sidecar binds 127.0.0.1:11435 with OLLAMA_NUM_GPU=0 (CPU). Measure RSS (/api/ps size), not VRAM.

If Compose cannot allocate a project network (Docker IPAM pool exhausted), the same sidecar is:

docker run -d --name 2dph-reasoner \
  -e OLLAMA_NUM_GPU=0 \
  -p 127.0.0.1:11435:11434 \
  -v 2dph-reasoner-ollama:/root/.ollama \
  ollama/ollama:latest

Real Hugging Face ids

Role	HF id	Ollama tag (this bake-off)
RAM / 9B	Qwen/Qwen3.5-9B	qwen3.5:9b
Quality 27B (CPU)	prism-ml/Bonsai-27B-gguf (derived from Qwen3.6-27B)	MichelRosselli/bonsai-27b:Q1_0
Quality 27B (full)	Qwen/Qwen3.6-27B	not pulled on this CPU box

There is no official Qwen3.6-9B. Do not invent that id.

Qwen3.5-9B has documented upstream tool-call XML bugs. A 9B win on tools is only claimed if this bake-off records OpenAI tool_calls (not <tool_call> XML in content).

The 2dph API image does not COPY GGUF/safetensors. Pull at runtime into the reasoner-ollama volume.

Live CPU run

Host sidecar: Ollama 0.32.9, OLLAMA_NUM_GPU=0, 127.0.0.1:11435, device: cpu, vram_mb: 0. Date: 2026-08-13. Same three prompts (search / get / audit). PicoClaw binary was not used; the OpenAI tools payload is the surface it would send.

Model	HF id	tool_call	xml_leak	rss_mb	latency_ms (search/get/audit)
qwen3.5:9b	Qwen/Qwen3.5-9B	3/3	0	5790	50059 / 51678 / 31980
MichelRosselli/bonsai-27b:Q1_0	prism-ml/Bonsai-27B-gguf	3/3	0	21951	327702 / 166830 / 119808
Qwen/Qwen3.6-27B	Qwen/Qwen3.6-27B	not loaded	—	—	too heavy for this CPU box

Both loaded models emitted OpenAI tool_calls (not <tool_call> XML) on this runtime. Do not claim 9B is better at tools — the score is tied at 3/3. 9B is smaller and faster. Bonsai RSS includes weights + KV (size from /api/ps); first Bonsai prompt includes cold load.

Re-run:

REASONER_BASE_URL=http://127.0.0.1:11435/v1 REASONER_MODEL=qwen3.5:9b \
  ./bin/reasoner/bakeoff.go --json
REASONER_MODEL=MichelRosselli/bonsai-27b:Q1_0 ./bin/reasoner/bakeoff.go --json