SIA is an open-source Self Improving AI framework with 2,100+ stars on GitHub that autonomously improves the performance of an AI system — a model or an agent — on a benchmark task, iterating on the system itself rather than asking you to hand-tune it.
Most AI development is a manual loop: run an evaluation, inspect the failures, adjust prompts, weights or agent logic, and repeat. SIA turns that loop into the product. You point the framework at an AI system and a benchmark that defines success, and SIA runs cycles of evaluation and modification, keeping the changes that measurably improve the score. That makes it relevant both to model-centric work, where the thing being improved might be a fine-tune or configuration, and to agent-centric work, where prompts, tools and workflows are the levers. The project comes from Hexo Labs and is published under the MIT license, and it is under active development with updates continuing through 2026. As with any automated optimisation, the quality of the benchmark matters: SIA improves whatever you measure, so it rewards teams who can express “better” as a concrete, repeatable evaluation.
SIA is free and open source under the MIT license. Running it still costs whatever the underlying model APIs and compute for your benchmark evaluations cost, but the framework itself has no paid tier.
AI engineers and researchers who maintain an agent or model pipeline against a measurable benchmark and want to automate the tedious evaluate-tweak-repeat cycle instead of doing every iteration by hand.