allenai/SERA-8B-GA

SERA-8B-GA: An Open-Source Coding Agent

SERA-8B-GA, developed by the Allen Institute for AI (Ai2), is an 8 billion parameter open-source coding agent. It is the fourth model in Ai2's Open Coding Agents series, built upon the Qwen 3-8B base model and fine-tuned with GLM-4.5-Air (110B) as the teacher model. This model is specifically designed for automated software engineering tasks, leveraging a 32K token context length.

Key Capabilities

• High Performance on SWE-bench: Achieves a resolve rate of 31.7% on SWE-bench Verified, outperforming other 8B open-source models like SkyRL-8B (9.4%) and Nex-N1-8B (20.3%).
• Automated Software Engineering: Excels at tasks such as bug fixes, feature implementation, and code refactoring.
• Synthetic Data Training: Trained on 200,000 synthetic coding agent trajectories generated using Soft Verified Generation (SVG), a two-rollout pipeline that removes the need for test infrastructure.
• CLI Integration: Easily usable via the sera CLI for seamless integration and deployment.

Good For

• Automated software development: Ideal for automating common coding tasks in Python repositories.
• Repository specialization: Can be fine-tuned on private codebases to create highly specialized coding agents.
• Research: Suitable for studying coding agents, data generation methods, and agent behavior.

Limitations

• Primarily validated on SWE-bench Verified (Python repositories); performance on other languages or benchmarks is unknown.
• Performance is largely bounded by the GLM-4.5-Air teacher model's capabilities.
• May generate insecure or incorrect code; all outputs require human review and testing.