This is a standalone fine-tuned version of Gemma 4 12B for software project and code review, converted to native 4-bit MLX for Apple Silicon. The fine-tuned weights are merged into the model; no separate adapter is required.

It is designed to produce focused, actionable review findings, use repository tools effectively, and recommend clear, self-explanatory implementations.

The release uses a calibrated merge of the fine-tuned and base weights to preserve general instruction-following quality while retaining the reviewer behavior.

Download it from the model's public Hugging Face URL in LM Studio, or run:

For typical local review work, a 16,384-token context is a practical starting point.

The release candidate scored 97/100 on a balanced comment-and-naming policy suite: 50/50 required-change cases and 47/50 clean controls. It also returned NO_COMMENT for all 10 clean changes in a separate held-out review sample.

The focused suite evaluates one policy violation at a time. When a small diff contains both an authored comment and an opaque name, the model can return only one of the two findings; use separate policy passes when exhaustive style enforcement is required.