Mava is a research-friendly codebase for fast multi-agent reinforcement learning experimentation in JAX, from InstaDeep, with 938 stars on GitHub.
Mava lets researchers experiment with multi-agent reinforcement learning (MARL) at very high speed. Its implementations are deliberately single-file JAX programs built for rapid research iteration: you can read an entire algorithm in one place, hack it, modify it and test new ideas fast, instead of untangling a framework. Under the hood, Mava’s distributed architecture and its state-of-the-art algorithms — including the Sable algorithm family — scale seamlessly across devices, from a single accelerator to larger clusters. It comes from the research team at InstaDeep, a leading AI research company, and is accompanied by a published paper describing the system design, making it as much a reference implementation as a toolkit. Environments and baselines are included so researchers can reproduce published results before changing anything, and the codebase favours readability — each algorithm’s core logic stays in a single file you can read end to end. That transparency is deliberate: Mava is as much a teaching artefact for how modern MARL systems are engineered as it is an experimentation tool for new research.
Mava is free and open source under the Apache-2.0 license. There are no fees or tiers; your only costs are the compute you run your experiments on.
Mava is for reinforcement learning researchers and engineers who work on multi-agent problems — cooperative or competitive agents, simulation environments and MARL algorithm development. If you want a codebase you can fully understand and bend to a new research idea within a day, Mava’s single-file design is its main attraction.