AI-Hypercomputer/maxtext

Health v2 · maintenance only

Adoption not part of the score

2408 stars · 592 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

MaxText is a high-performance, scalable open-source LLM training library written in pure Python/JAX, targeting Google Cloud TPUs and GPUs. It provides reference implementations of models like Gemma, Llama, DeepSeek, Qwen, and Mistral with support for pre-training at massive scale and post-training techniques such as SFT, GRPO, and GSPO.

Use cases

• pre-train large language models on TPU clusters
• fine-tune Llama or Gemma with supervised fine-tuning
• run reinforcement learning post-training like GRPO for LLMs
• scale LLM training to tens of thousands of chips
• train DeepSeek or Mistral models in JAX
• benchmark high MFU LLM training on GPUs and TPUs

When to choose

• you need to train or fine-tune LLMs on Google Cloud TPUs or GPUs at large scale
• you want a pure Python/JAX codebase that is easy to fork and modify
• you need high hardware utilization without hand-written kernel optimizations
• you want reference implementations of popular open model architectures

When to avoid

• you need PyTorch-based training or ecosystem compatibility
• you only need LLM inference rather than training
• you lack access to TPU or GPU hardware
• you need diffusion or image generation models (see MaxDiffusion instead)

Facets

library · maturity active

llm-training machine-learning deep-learning large-language-models deep-learning machine-learning python cloud jax tpu fine-tuning pre-training sft grpo gemma llama deepseek mistral xla gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="AI-Hypercomputer/maxtext")