SERA-8B: An Open-Source Coding Agent
SERA-8B, developed by the Allen Institute for AI (Ai2), is an 8 billion parameter model designed as a state-of-the-art open-source coding agent. It is the third model in Ai2's Open Coding Agents series, built on the Qwen 3-8B base model and trained using GLM-4.6 as a teacher model.
Key Capabilities & Performance
- Automated Software Engineering: SERA-8B is optimized for tasks such as bug fixes, feature implementation, and code refactoring.
- SWE-bench Verified: Achieves a 31.7% resolve rate on the SWE-bench Verified benchmark at a 32K context length, outperforming other open-source models like SkyRL-8B and Nex-N1-8B in its size class.
- Training Method: Utilizes supervised fine-tuning on 200,000 synthetic agent trajectories generated via Soft Verified Generation (SVG), a two-rollout pipeline that removes the need for test infrastructure.
Intended Use Cases
- Automated Software Engineering: Ideal for automating various coding tasks within a development workflow.
- Repository Specialization: Can be fine-tuned on private codebases to create highly specialized coding agents.
- Research: Valuable for studying coding agents, data generation methods, and agent behavior.
Limitations
- Primarily validated on SWE-bench Verified (Python repositories); performance on other languages is unknown.
- May generate code with security vulnerabilities or inaccuracies, requiring human review and testing.
- Inherits biases from its base and teacher models and lacks safety filtering.