bend-heads · support map (no weights yet)

One model per size. Heads swap on the same loaded copy. A head is not a model: it is the small interface that turns a loaded base model into a mode (structured decisions of a given family). Pick a size model (bend-4b, …), pick a head, switch modes without reloading anything. This repo holds no weights yet; it is the map.

The heads

• kev-native: ships with bend-4b; the base's own decision mode. Chat + decisions from one load. Status: native. Kev on the phone/Mac engine: the same decision request, 12 runs, byte-identical. A Kev session mixing chat and decisions has not been gated yet.
• clef-flash: Cloudflare's Clef family. Runs end-to-end on the Bend engine (parity |Δp| 0.0092 vs reference after one wire fix). Status: wired; GPU path int8 ≤0.013, top-1 24/24; MLX path runs. Clef Flash 9B · RTX 3090 · SYS1-DUAL: 28/28 decisions byte-identical over 24 goldens with chat in the same process; one process, 6,491 vs 12,597 MiB card memory vs two resident engines.
• jev-r2: TypeSafe's Jev family. 313/313 byte-precise runs (max |dp| 2.2e-6). Status: wired.
• clef-omni: Cloudflare's omni release (2026-10-09). Status: not yet run here; in queue.

Results on one head are not results for another; each figure above is labelled with the model it was measured on.

The rule

A head in the wired shape (a trained interface over a supported base, Qwen-family bases included) is the shape the runtime already loads. A genuinely new head architecture needs a port. The parity receipts above are the gate: a head is "listed" when its runs are checkable.

See SUPPORT-MAP.md for the source-labeled map.