SERA (open coding agents 8B, 14B, 32B)

2026-01-27 · Publisher : Ai2 · Domain : Artificial Intelligence

Ai2 has released SERA, a family of open coding agents (8B, 14B, 32B, Qwen 3 base) with weights, code, synthetic data and a CLI, trained at low cost thanks to soft-verified data generation.

### Context High-performing coding agents generally rely on training data that is costly to verify. Ai2 proposes a synthetic generation method where trajectories do not need to be fully correct, and publishes the recipe. ### What the model brings - "Soft-verified generation": training on partially correct data avoids costly full verification while keeping effectiveness, according to Ai2. - SERA-32B solves 54.2% of SWE-Bench Verified at 64K context and 49.5% at 32K. - Ai2 considers it competitive with Devstral Small 2 and GLM-4.5-Air under comparable settings. - A stated cost of about $400 to reproduce the best previous open results and about $12,000 to match Devstral Small 2, or about 40 GPU-days on 2 Hopper GPUs. - Per-repository specialization: 52.23% on Django and 51.11% on SymPy. ### Architecture and specifications 8B, 14B and 32B models on a Qwen 3 base, trained up to 32K context (SERA-32B also evaluated at 64K and beyond). License: Apache 2.0 (32B HF model card). Openness: weights, training code and recipes, synthetic datasets in a model-independent format, the SERA CLI command-line tool (PyPI) with Claude Code integration. The pretraining of the Qwen 3 base is not open: an open post-training recipe, not a fully open model. ### Cost and access | Item | Detail | |---|---| | Official API | None; CLI and open weights | | Training hardware | About 40 GPU-days on 2 Hopper GPUs to reproduce the open state of the art | | Weights | SERA-32B, 8B (January 27, 2026), 14B (February 3, 2026) | ### Benchmarks | SWE-Bench Verified (Ai2) | Resolution rate | |---|---| | SERA-32B, 64K context | 54.2% | | SERA-32B, 32K context | 49.5% | Source: Ai2 blog post of January 27, 2026, Ai2 measurements; the comparisons with Devstral Small 2 and GLM-4.5-Air (open) are qualitative in the excerpt consulted. ### Versus closed models No closed model is compared in the sources consulted. The positioning is that of a reproducible, low-cost recipe rather than a frontier model: results are compared with open coding models (Devstral Small 2, GLM-4.5-Air). A score of 54.2% on SWE-Bench Verified is measured by Ai2 and is not linked here to a closed-model score from the same period. The advantage for a company is being able to specialize a small agent on its own repository at a stated cost of a few hundred to a few thousand dollars. This addresses a code sovereignty need that closed API models do not cover. ### Key takeaways SERA makes the training of specialized coding agents accessible. The cost and score figures are Ai2's.

Tags : LLM, Open Source, Poids, code et données ouverts, Ai2, SERA, Agents de code, Apache 2.0