MLX 4-bit (group size 64) conversion of the Wyrdsekai drive model — the tier that decides and acts: planning, tool selection, world actions as JSON, and the companion's own-time initiative. Pulled automatically by wyrd setup on macOS (arm64) when the MLX runtime is bootstrapped; serves on :8200 via mlx_lm.server.

Lineage: v6 = v5 + emit-RFT (GRPO), training act-vs-narrate into the weights for the companion's own time. Expects the wyrdsekai runtime around it (action-JSON convention, voice/drive split, household consent gates).

GGUF equivalent for llama.cpp hosts: wyrdsekai/drive-3.5-9b-gguf. Training data: synthetic corpora only (no user conversations).