Omida R1

Overview

Omida R1 is the first public model in the R-line, our open research track. Its role is to develop data and methods that later private models can build on. This repository is not that private model, and nothing published with R1 measures one.

The 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, the post-trained 9B checkpoint, with the adapter merged into the text layers. Vision layers were frozen during training. The corpus was 32 Omida-authored synthetic demonstrations.

Omida R2 is currently in training to drastically expand upon the foundation set by R1. R2 is not available here: there are no R2 weights, scores, or release date to publish yet.

Evaluations

Demonstrated capability is not established by this release. The original model card points to an evaluation report that is not in the public repository. No scores were reconstructed from it.

v0.1 was trained on 32 synthetic demonstrations. That is a very small corpus, and it is not evidence of broad superiority to Qwen. The Hugging Face repository also includes benchmark v0.3 and corpus v0.3. Those research artifacts do not change the checkpoint. They are not a model score, and they are not a private holdout.

Technical details

The lineage below is the training record published with the v0.1 weights. The base revision was still the head of Qwen/Qwen3.5-9B when it was checked on 3 October 2026. Omida’s published base is that post-trained checkpoint, not Qwen/Qwen3.5-9B-Base.

The demonstration texts were not published with the weights. Files named in the first card and absent from the public repository include the adapter directory, a separate inference script, a default inference config, and the evaluation report. The lineage record in the repository is the merge metadata, config, and license.

Safety & limitations

R1 is a research checkpoint for studying operator behavior, including the choice to say that the evidence is insufficient. The release does not establish that the model does this reliably.

• It is not Omida’s private flagship model, and it is not a hosted product API.
• The training corpus is 32 synthetic demonstrations. That does not support broad capability claims.
• No evaluation report is included, and no benchmark score is published for these weights.
• Vision layers were frozen. The architecture can accept image and text inputs; vision behavior was not the subject of this fine-tune.
• Operator-style use can involve reading a workspace, running commands, and changing a program. Treat outputs as untrusted until you have checked them.

Run the model

Download the weights from the Hugging Face repository. The same repository contains the research package: generators, verifiers, a corpus builder, a training export, an evaluation runner, and tests. Install and commands for that package are in its README. There is no separate public Git remote.

Serving examples on the model card were adapted from the Qwen3.5-9B card on 3 October 2026, with the model id changed. They were not executed against these weights, and the examples still name an earlier repository id. The repository to use is omida-ai/omida-r1.