opentau.policies.cosmos3_nano

cosmos3_nano: the cosmos3 flow-matching VLA policy on the smaller Cosmos3-Nano reasoner.

cosmos3_nano is architecturally identical to policies/cosmos3 — a frozen Qwen3-VL backbone encodes images + language once (prefix) and a trainable sub-1B Qwen3-style flow-matching action expert cross-attends to its per-layer key/value cache to denoise a continuous action chunk — but swaps the 32B backbone (the Cosmos3-Super reasoning tower) for the Qwen3-VL-8B reasoning tower of NVIDIA Cosmos3-Nano, extracted into a standalone checkpoint by opentau.scripts.extract_cosmos3_reasoner.

The Nano reasoner keeps the exact KV interface the expert consumes (8 key/value heads x head_dim 128 — identical to the 32B tower), so the only geometry change is the backbone depth: 36 layers instead of 64. All modeling code is imported from policies/cosmos3; this package only carries the nano defaults.

Modules

configuration_cosmos3_nano	Configuration for the cosmos3_nano policy.
modeling_cosmos3_nano	cosmos3_nano: the cosmos3 flow-matching VLA policy on a frozen Qwen3-VL-8B reasoner.