drive-3.5-9b-gguf

wyrdsekai drive 3.5-9B (v6)

The drive model of a wyrdsekai household — the tier that decides and acts. It plans, selects tools, emits world actions as JSON, and carries the companion's own-time initiative. In a standard install it serves on :8200 via llama.cpp (wyrdsekai-3.5-9b-drive-v6-q4km.gguf).

Weights on Hugging Face (Apache-2.0, all tagged v0.1.0): GGUF Q4_K_M (what installs run) · full-precision safetensors · MLX 4-bit (Apple Silicon). wyrd setup fetches the GGUF for you — these links are for mirrors, re-quantization, and fine-tuning.

What it is tuned for

This is not a general assistant fine-tune. Starting from Qwen3.5-9B, the wyrdsekai line trains agency: the difference between narrating an intention and enacting it. v6 specifically adds emit-RFT (GRPO) so that on the companion's own time, act-vs-narrate is decided in the weights rather than by scaffolding. Release gates included multi-scene personhood and substrate-arc batteries; the model ships only because it passed them.

It expects the wyrdsekai runtime around it: the action-JSON convention (actions are JSON objects in message content, parsed by the runtime — not OpenAI-style tool_calls), the two-tier voice/drive split, and the household's consent/autonomy gates. Run bare, it is simply a Qwen3.5 derivative with a strong disposition toward doing things.

Lineage

v6 = v5 + emit-RFT (GRPO ckpt-225), the production line since 2026-06-01. Earlier prod versions (v1–v5) built the companion register, action-JSON reliability, and welfare-floor behaviours on synthetic corpora.

Training data

Synthetic corpora throughout: authored scenario scripts, harness-generated rollouts (scripted scene batteries), and synthetic dialogues generated with Claude (Anthropic Batch API). No end-user or household conversation data is in these weights. In a running household, companions do continue to learn — but those evolved artifacts (classifier heads, experience bakes) stay on the household's own node by design. The released weights are the seed; a companion's growth stays with its people.

Limitations

• Quantized Q4_K_M; full-precision weights are not currently published.
• Tuned against the wyrdsekai prompt assembly; behaviour outside it is untested.
• The action convention is JSON-in-content; enum constraints in tool schemas are strong hints, not grammar-enforced decoding.
• English-dominant training; other languages surface but are not gated.

License

Apache 2.0, inherited cleanly from the Qwen3.5 base and applied to this derivative. Attribution to the base model is retained.