SERA uses a simplified training method called “Soft-verified Generation,” which does not require fully correct code examples. According to Ai2, training requires just 40 GPU-days and costs between $400 to match earlier open-source results and $12,000 to reach performance comparable to leading industry models. This makes training on proprietary code data feasible even for small teams. Technical details are available in the project blog.

32K context, Allen AI
32K context, Allen AI 

The models are compatible with Claude Code and can be launched using just two lines of code, according to Allen AI. All models, source code, and documentation are available on Hugging Face under the Apache 2.0 license.