Yandex says its voluntary internal AI reviewer handles about 4,000 pull requests daily, using Qwen 3.6 to find issues and MiniMax to filter them.
Yandex Tech says NeuroReview now examines about 4,000 employee pull requests each day, equivalent to roughly 40% of PRs created in Arcadia, the company’s monorepository. Use is voluntary and limited to internal code. The current production setup uses Qwen 3.6 27B to identify potential defects and MiniMax-M2-7 to validate the resulting comments.
The service runs on Yandex’s Comrade platform for LLM agents, with Temporal orchestrating the workflow. A generator reads the diff and relevant files to assemble candidate issues. A second agent receives those findings and the original changes, removes duplicates, and rejects comments it considers debatable or insufficiently supported. Before publication, the service also checks the cited file, line and code fragment against the repository.
Yandex evaluates configurations through A/B experiments on live PR traffic. Its main metric, Time Saved per PR, estimates engineering time saved when suggestions are accepted or issues are closed, while penalizing rejected issues. The company cautions that TSPR is a proxy for developers’ observable reactions, not a direct measurement of working hours, and uses it to compare configurations within the same product environment.
According to Yandex, the tested Opus configuration produced a TSPR about 30% higher than the Qwen configuration, but each check cost $3.42 rather than $0.041—approximately 83 times more. Running Opus once across 4,000 PRs would therefore cost about $13,700 per day, before repeat checks of revised diffs. Yandex selected Qwen for production as its preferred balance between quality and expense.
NeuroReview posts its first comment an average of two to three minutes after a PR is published, Yandex says; this interval is the time to initial feedback, not the duration of processing. A human reviewer often joins several hours later. Developers accept about 30% of the service’s comments. Around half of accepted findings concern business logic and another 16% are language-level bugs. Final decisions and responsibility for the code remain with people.
The service is strongest on local defects that can be supported by the diff and nearby code. It has more difficulty with architectural issues, interactions among multiple services and undocumented business rules. Context quality also deteriorates in very large PRs spanning hundreds of files, which can cause missed issues or additional noise. A clean result does not prove the code is correct, and NeuroReview does not perform architectural review.
The two-stage design illustrates a practical division of labor: one model can search broadly while another controls what reaches developers. The reported comparisons should not, however, be generalized to unrelated repositories or treated as isolated model benchmarks. Yandex tested complete production configurations—including models, prompts, tools, checks and the validator—and did not measure each component’s contribution separately.
| Generator | TSPR comparison | Cost per check | Comment |
|---|---|---|---|
| MiniMax-M2-7 | Baseline | Not stated | Initial configuration; MiniMax was also used as the validator |
| Gemini | +37.22% versus baseline | Higher than Qwen; amount not stated | The source does not identify the model version |
| Qwen 3.6 27B | +62.77% versus baseline | $0.041 | Selected for the production pipeline |
| Opus | About 30% higher than Qwen | $3.42 | Approximately 83 times more expensive than Qwen |
Sources
Originally published by AI Feed: Yandex Details Qwen-Based NeuroReview for Internal Code Checks.
Top comments (0)