MaxText

MaxText is Google’s open-source library for training large language models in JAX: simple, high-performance and built to scale from a single machine to thousands of chips. It has around 2,400 GitHub stars.

Overview

Developed by Google’s AI Hypercomputer team, MaxText provides reference implementations and training pipelines for the major open model families — Gemma, Llama, Mistral and Mixtral, Qwen and DeepSeek among them — covering pre-training and post-training such as SFT. It is written in pure Python on JAX, leans on Flax for modelling, and is heavily optimised for TPUs while also running on GPUs. Google used MaxText in training its own open models, so its recipes are a practical starting point for reproducing or extending them, and support for new open models lands here quickly after release. The codebase doubles as a performance reference: its configs show how to shard models and data to reach high hardware utilisation at scale. MaxText is under very active development, with commits landing in 2026, and its documentation is extensive.

Key Features

• Pre-training and post-training pipelines for major open LLMs
• Reference configs for Gemma, Llama, Mistral, Qwen, DeepSeek
• Scales to thousands of TPU or GPU chips with JAX sharding
• Pure-Python JAX codebase that is easy to read and fork
• Optimised kernels and checkpointing for large runs
• Detailed documentation and model-specific guides

💳 Pricing

MaxText is free and open source under the Apache-2.0 license, and there is no paid edition. You pay only for the TPU or GPU capacity you train on.

Who Is It For

ML platform teams and researchers training or fine-tuning open LLMs at serious scale, especially on Google Cloud TPUs, and JAX users who want LLM training without a PyTorch detour.