MiniMax's blockbuster model goes open source! Huawei Ascend will integrate it, processing over 100 trillion tokens in 3 days

What happened: MiniMax released the open weights of its M3.1 Flash model, a 428 billion parameter model with about 23 billion parameters activated per token. Why it matters: MiniMax said it is working with Huawei Ascend, T-Head, AMD, and platforms such as Fireworks on first day integration, while Kunlunxin completed Day 0 adaptation and Nvidia NeMo, vLLM, SGLang, and Together AI announced support. Open weights let developers download and self deploy from Hugging Face and ModelScope.

MiniMax released the open weights of its M3.1 Flash model, a 428 billion parameter model with about 23 billion parameters activated per token.

MiniMax said it is working with Huawei Ascend, T-Head, AMD, and platforms such as Fireworks on first day integration, while Kunlunxin completed Day 0 adaptation and Nvidia NeMo, vLLM, SGLang, and Together AI announced support. Open weights let developers download and self deploy from Hugging Face and ModelScope.