rumik-oss 1

blog | model card

Private test deployment. The checkpoint runs directly on Hugging Face ZeroGPU, including autoregressive generation and Mimi decoding. No external inference API is used.

The checkpoint is loaded from the private rumik-ai/rumik-oss-1 repository at revision e4f56800cd96e6234ddc0494779a248277a5e141. Weights, tokenizer, codec, and inference implementation are copied from the verified ayaserve bundle without another adapter merge. The original sampling path is retained: temperature 0.8, top-k 30, sampling enabled, and 2,048 output tokens by default. Top-p filtering is not implemented in that runtime.

The Tiny Aya-derived weights retain their upstream non-commercial license and acceptable-use terms. Mimi retains its CC-BY-4.0 license. See the bundled model card and the model release.

Deployment requires a Space secret named HF_TOKEN with read access to the private checkpoint. Select ZeroGPU hardware (zero-a10g). No ayaserve URL or API key is used. Requests are serialized and reserve up to 120 GPU seconds.