Omida R1 is a research checkpoint
Omida R1 is an open-weight model for research on operator behavior: reading a local workspace, running commands, repairing a small program, and stating a conclusion only when the evidence supports it. The R-line is the public research track. Its role is to develop data and methods that later private models can build on. This release is not that private model, and nothing here measures one.
The published weights are the v0.1 merged checkpoint. R1 was not trained from scratch. It is a QLoRA supervised fine-tune of Qwen/Qwen3.5-9B, with the adapter merged into the text layers. Vision layers were frozen. The training corpus was 32 synthetic demonstrations. The license is the upstream Apache-2.0 text.
What this release includes
- v0.1 merged weights on Hugging Face
- The upstream Apache-2.0 license
- A lineage record: model config and merge metadata
- A research package in the same repository, with generators, verifiers, a corpus builder, an evaluation runner, and tests
- Benchmark v0.3 and corpus v0.3, which do not change the checkpoint and are not a model score
What it does not establish
Demonstrated capability is not established by published scores. The evaluation report named in the original card is not in the repository, and no scores were reconstructed. A corpus of 32 synthetic demonstrations is not evidence of broad superiority to Qwen. Serving examples on the model card were adapted from Qwen and were not run against these weights.