MENLO PARK, Calif. — Corepass, which makes the Remy coding agent, has finished training the first model in its own Scout line, and the company’s measurements suggest it finds the right code more accurately and faster than the model it started from.
The model belongs to a line Corepass calls Scout — lightweight specialists, in the company’s description, aimed at code search and tasks whose results can be verified. Scout’s first model, at 4 billion parameters, has completed reinforcement learning on real codebases and is integrated into code-search workflows, the company says.
What the numbers show
Corepass measured the model on all 848 tasks of SWE-Explore-Bench, a code-exploration benchmark. Its retrieval F1 — a standard score that balances how many of the relevant files a system finds against how many irrelevant ones it returns — was 0.180, about 2.3 times the base model’s 0.077. It also worked faster, taking 5.4 seconds per task against the base model’s 7.6, a reduction of 29 percent.
| Measure | Scout (4B) | Base model | Change |
|---|---|---|---|
| Retrieval F1 | 0.180 | 0.077 | 2.3× |
| Time per task | 5.4 s | 7.6 s | −29% |
The figures come from the company’s own report and have not been replicated by an outside group. An F1 of 0.180 also leaves considerable room to improve in absolute terms; the gain is measured against the starting model, not against other systems.
Scout’s first model more than doubled its base model’s retrieval score, by the company’s own measure.
Two larger models in development
Scout is the smallest of three model lines. Voyager, described as balanced in capability, and Horizon, the company’s planned flagship for reasoning, are both in development. Corepass says they are supported by two systems it calls Vera and Forge, which handle real-world tasks, evaluation and training. The company says its model research centers on post-training, alongside continued pre-training work.
Not yet open to outside developers
The models are not available through a public interface. Corepass says external API access is not yet open, and that model access, agent execution and related tools will be opened in stages through the Corepass Platform.