AI2 launches Open Coding Agents, drastically reducing the cost to create programming agents adapted to any repository

The Allen Institute for AI (AI2) announced the launch of Open Coding Agents, a new family of open programming agents that promises to democratize the development of agents capable of generating, reviewing, debugging, and maintaining code in any repository, including private codebases. Unlike dominant closed models on the market, AI2's proposal combines open-weight models with a highly efficient and reproducible training method, drastically reducing costs and technical barriers. The main advance lies in the post-training method, which allows creation of high-quality synthetic data directly from real repositories, without depending on expensive pipelines or complex test infrastructure. With this approach, AI2 demonstrates that it is possible to reproduce the performance of the previous best open-source model for about $400 in compute, or achieve performance comparable to industry-leading models with approximately $12,000, making agent specialization viable for small teams and independent labs. The first model in the family is SERA (Soft-verified Efficient Repository Agents). SERA-32B achieves 54.2% resolution on the SWE-Bench Verified benchmark in a 64K token context, outperforming previous open-source models of similar size and rivaling much larger commercial agents. Technically, the differentiator is the use of synthetic generation with soft verification, where partially correct patches are still useful for teaching agentic behavior, eliminating the need for rigid and costly validation. Another key point is the ability to specialize in private code. In tests with large repositories such as Django and SymPy, SERA models fine-tuned with just 8,000 synthetic trajectories outperformed teacher agents with over 100 billion parameters. This proves that smaller models, when well adapted to the real context of a repository, can match or exceed larger generalist systems, with lower operational cost, lower latency, and greater control. In terms of infrastructure, AI2 worked with NVIDIA to optimize inference on H100 and Blackwell GPUs, achieving up to 8,600 tokens per second in NVFP4 precision. Furthermore, the entire ecosystem—models, data, code, and training recipes—is fully open, allowing developers to launch production-ready agents or perform fine-tuning with few command lines, without prior LLM training experience. The launch of Open Coding Agents signals an important shift in the AI-for-software market: programming agents cease to be exclusive to large companies and become accessible strategic assets.