MiniMax launches M3 with native multimodality and a 1M-token context

MiniMax released the open-weight M3 on June 1, 2026. Its MiniMax Sparse Attention architecture delivers 9× faster prefill and 15× faster decoding than the previous generation at a 1M-token context. M3 supports text, images, and video natively. It scored 59.0% on SWE-bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro, and 70.06% on OSWorld-Verified. API access was available at launch, with weights and a report planned for Hugging Face within about ten days.

Visit original source