Rixven

Kerven Vega 1

A reviewer for the code your other model wrote.

Developers now spend more hours reading AI-generated code than writing their own — and the bugs that survive review are exactly the ones that read fluently. Kerven Vega 1's only job is catching those.

A row of abstract code blocks with a single block circled in ember orange, as if flagged by a reviewer — illustrating Kerven Vega 1's single job of catching the one wrong line. loading=eager fetchpriority=auto />

What it's trained to catch

Hallucinated APIs

Calls to functions, methods, or endpoints that do not exist in the library being used — plausible names, no such symbol.

Wrong signatures

Real functions called with the wrong argument count, order, or type — the kind of mismatch that only surfaces at runtime.

Confident logic bugs

Code that reads correctly and does something the comment above it does not describe.

Why only two languages

The base model already reads more than forty languages. That was never the constraint. A reviewer is only as trustworthy as the examples it was corrected on, and correcting it carefully, with real injected bugs and verifiable labels, is what's slow.

Python and JavaScript first because that's where the most AI-generated code is already shipping. Every other language reuses the base model's existing knowledge — it's an expansion, not a rebuild, once the reviewing skill itself is proven.

How it's built

Base model
Qwen3-Coder-30B-A3B — 30B total, ~3B active (MoE), Apache 2.0
Languages, v1
Python and JavaScript only
Training
Supervised fine-tuning, then GRPO against a verifiable bug-injection reward
Compute
JAX / Flax on a TPU v5e-8 (Google TPU Research Cloud)
Data
Synthetic — 20–30k examples of AI-written code with injected, labelled bugs
Release
Open weights on Hugging Face, plus a hosted API

After v1

In roughly the order we'll get to them.

  1. More languages — reusing the base model's existing coverage.
  2. Fix suggestions — not just flagging the bug, but proposing the correction.
  3. Doc-grounded review — retrieving real, current API references at inference time, to cut false positives.
  4. Repo context — reviewing a change against the codebase it lands in, not just the diff.
  5. CLI and CI integration — a GitHub Action, once the model itself has earned it.

Weights are open. Hosting is optional.

Run it yourself from Hugging Face, or use the hosted API when it's available. We'll announce the release here and in News.

Get notified at launch