AI startup MiniMax officially released the open-weight version of its flagship model, MiniMax M3, on Hugging Face today. The model uses a native multimodal Mixture-of-Experts (MoE) architecture with 428B total parameters and 128 experts, activating only 4 experts (about 23B parameters) per token. This design balances knowledge capacity and inference speed. Alongside the bfloat16 main version, MiniMax also published a MXFP8 quantized variant to reduce VRAM requirements for broader deployment.
428B MoE Architecture: Only 23B Active Parameters per Token
The 60-layer model achieves high capacity with low computation by routing each token to just 4 experts. Compared to dense models of similar size, M3 cuts inference cost significantly. The quantized version further lowers hardware barriers, enabling operation on consumer GPUs.
Proprietary MSA Mechanism: 15x Decoding Speedup at 1M Context
M3 supports a 1-million-token context window. The key enabler is the MSA (MiniMax Sparse Attention) mechanism, which uses a "lightning indexer" for efficient block-sparse attention. According to the technical paper, MSA achieves up to 9x prefill speedup and 15x decoding speedup in extreme long-context scenarios, dramatically lowering the compute cost of long-context AI.
Native Multimodal from Step Zero, Top Coding and Agent Performance
Rather than adding multimodal capabilities post-training, M3 was trained from scratch on text, images, and video. This endows it with strong long-video understanding and complex desktop operation skills. On coding benchmarks, M3 scored 59.0% on SWE-Bench Pro and 66.0% on Terminal Bench 2.1, making it suitable for multi-step reasoning and tool-calling agent workflows. It also offers "Thinking" and "Non-Thinking" modes for flexible trade-offs between deep reasoning and low latency.
Deployment Guidance: SGLang, vLLM, Optimized for NVIDIA Blackwell
M3 is available on Hugging Face and Unsloth. The official cookbook recommends deploying with SGLang, vLLM, or Transformers (with trust_remote_code=True). The model is optimized for NVIDIA Blackwell platforms, enabling developers to build next-gen multimodal agent applications at lower cost when combined with the MXFP8 quantized version.