SERA

AI2 has introduced Open Coding Agents, a set of powerful open-source coding models and a training method. This allows anyone to build custom coding agents for any software project. These agents can help with tasks like writing new code, reviewing existing code, fixing bugs, maintaining code, and explaining what code does.

Previously, using coding models with private code was difficult because they didn't know about that specific code. Training models on private data was also expensive and hard because it was challenging to create the right kind of training information. AI2's new method makes this much cheaper and easier, so even small teams can train their own agents.

SERA (Soft-verified Efficient Repository Agents) is the first product in this new family. The strongest model, SERA-32B, performs very well on coding challenges, doing better than other open-source models of similar size. It also needs much less computing power to train. SERA models are made to work well with Claude Code and can be trained on private codebases without costing a lot. NVIDIA has helped make SERA run faster on their hardware.

Everything released, including the models, the Claude Code connection, and the training instructions, is open-source. This means researchers can look at it, build on it, and get the same results. A key new idea is that these agents can learn from private data. Even smaller, open models can perform as well as or better than larger models when trained on specific codebases. For example, SERA-32B can perform better than its larger teacher model on projects like Django and Sympy after being trained on a small amount of data at a low cost.

Benefits

SERA makes it much easier and cheaper to train coding agents that understand specific private codebases. It uses a new training method that creates training data more efficiently, reducing the need for extensive testing. This allows organizations to customize small models affordably, replacing complicated training processes with simpler ones.

Use Cases

SERA is designed for practical coding tasks. Developers, researchers, and small teams can use it for code generation, code review, debugging, and code explanation. It can be used locally, in the cloud, or fine-tuned on private codebases. SERA models can be adapted to understand specific projects like Django, SymPy, and Sphinx, improving their performance on those particular codebases.

Pricing (ONLY include if available)

The cost to reproduce strong open-source results with SERA is approximately $400, which is significantly less than many other methods. Reproducing top open-weight models typically costs around $12,000.

Vibes (ONLY include if available)

Specialized SERA models, trained on specific code repositories, have consistently performed as well as or better than larger teacher models. This shows that a smaller, fine-tuned model can offer similar performance with lower costs and less operational effort. The training process is now competitive with large, complex projects, making agentic coding progress more accessible and repeatable.

Additional Information (ONLY include if available)

SERA models range from 8 billion to 32 billion parameters and are built on Qwen 3, trained for up to 32,000 context length. An update introduced SERA-14B, a new 14-billion-parameter model, and updated training datasets in a format that makes them easier to filter and analyze. The entire release package includes models, code, generated agent data, and a full recipe for data generation, all aimed at making it easy to reproduce and customize.