Developers who want an open coding model tuned on their own repo, cheaply; free weights, and it works with Claude Code.

Kenning's take

What it does

SERA (Soft-verified Efficient Repository Agents) is a family of open coding models from Ai2, from 8 billion to 32 billion parameters, built on Qwen 3. Ai2 also released the training method, which it says can match the previous best open-source result for about $400 of compute, so teams can tune a model on their own code. The models are on Hugging Face, with a SERA CLI on PyPI and a setup that works with Claude Code.

  • Open models from 8B to 32B parameters, plus SERA-14B
  • SERA-32B scores 54.2% on SWE-Bench Verified
  • Low-cost training recipe for private codebases
  • Works with Claude Code out of the box
  • Open training datasets

Who it suits. Developers and research teams who want an open coding agent tuned to their own code.

Key facts

Category
AI models
Made by
Ai2 (Allen Institute for AI)
Based in
United States
Released
27 January 2026
How to get it
download from Hugging Face; SERA CLI via PyPI (ai2-sera-cli) and GitHub
Works on
Hugging Face, CLI
In the UK
Yes
Free plan
free open weights

FAQ

When was SERA released?

Ai2 released SERA on 27 January 2026 and later added SERA-14B.

Is SERA free?

Yes, the models and training data are openly released; you pay for your own compute.

Does SERA work with Claude Code?

Ai2 says it is compatible with Claude Code out of the box and provides a setup script.

Facts from the maker's own pages and the sources listed. Checked 2026-10-11.