MaxText is a Python/JAX-based open-source LLM library focused on pre-training and post-training at scale, with multi-model support and a recent update cadence.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
MaxText is a high performance, highly scalable, open-source LLM library and reference implementation written in pure Python/JAX and targeting Google Cloud TPUs and GPUs for training. It provides a library of high performance models to choose from, including Gemma, Llama, DeepSeek, Qwen, and Mistral, with pre-training and scalable post-training support (SFT and RL methods).
How it works
MaxText leverages JAX/Amp tooling and integrations such as Flax, Tunix, Orbax, Optax, and Grain to enable pre-training and post-training workflows at scale. It supports models across multiple families (Gemma, Qwen, DeepSeek, Kimi, Llama, Mistral) and includes mechanisms for SFT and RL methods (GRPO, GSPO). The project emphasizes MFU and tokens/second efficiency across varying cluster sizes.
Getting started
Installation is described via an installation guide referenced in the README: "See our installation guide to install MaxText with pip from PyPI". The README also provides a decoupled mode guide: "Decoupled mode". Exact commands for setup are not provided in the truncated README text here, but the project points to online docs for getting started.
Recent releases
Latest releases include:
- maxtext-v0.2.3 (2026-06-12): Upgraded JAX to 0.10.0 for pre-training and 0.10.1 for post-training; new vLLM-Powered Evaluation Framework for running lm-eval and evalchemy.
- maxtext-v0.2.2 (2026-05-08): Upgraded JAX to 0.9.2; simplified APIs for accessing MaxText models; included maxtext_with_gepa.ipynb.
- maxtext-v0.2.1 (2026-03-23): Introduced maxtext[runner] installation option to build Docker images without cloning the repository; references to installation instructions.
- maxtext-v0.2.0 (2026-03-06): Qwen3-Next support; new tpu-post-train target in PyPI; notes on PyPI interactions.
- maxtext-tutorial-v1.5.0 (2025-12-30): No specific changes shown in excerpt.
Traction
Stars: 2379. Forks: 578. Open issues: 294. Language: Python. License: Apache-2.0. Created: 2023-02-28. Last push: 2026-08-04. The repository has activity across releases and a broad set of topics, including jax, llama, gemma, and fine-tuning, indicating active development.
Behind the repo
Not applicable from provided data.
Caveats
License: Apache-2.0. Created: 2023-02-28. Last push: 2026-08-04. The README references external installation guides and docs; exact installation commands are not included in the excerpt. README indicates Python 3.12 as primary supported version due to compatibility notes. No explicit age-related caveats beyond last push date provided here.