32B (based on Qwen 3-32B)
Apache 2.0
First release in Ai2's Open Coding Agents series, positioned as a specialized software-engineering agent rather than a general chat model. Official model-card reporting cites 49.5% ± 1.9% on SWE-bench Verified at 32K context (with related comparisons showing competitiveness against larger open-weight models). Training uses Soft Verified Generation (SVG), a two-rollout synthetic trajectory pipeline where a teacher model edits code, a synthetic PR description is produced, and a second rollout reproduces the change; card claims SVG reaches target quality at substantially lower cost (26x cheaper than RL and 57x cheaper than prior synthetic-data methods) with total generation+training cost around $2,000 (40 GPU-days). The released training set size is 25,000 synthetic coding-agent trajectories. Deployment guidance includes vLLM-based self-hosting and sera-cli integration with Claude Code, with references to Modal and endpoint-based operation. The release family also includes SERA-32B-GA (46.6%), SERA-8B (31.7%), and SERA-8B-GA (31.7%), with teacher-model differences noted in the card. Limitations section flags a known SWE-bench artifact where the model may attempt a nonexistent `submit` tool call in some workflows.