MiniMax Releases Open-Weight M3.1-Flash Multimodal MoE Model

MiniMax-M3.1-Flash is presented as a native multimodal mixture-of-experts model with about 428B total and 23B activated parameters, accepting text, image, and video inputs with up to a 1M-token context. It builds on MiniMax-M3, adding sparse attention throughout, quantization-aware training, and DSpark speculative decoding, with a draft head included. The model card reports benchmark results for agentic coding, computer use, and cybersecurity tasks against several other models.

Key takeaways

• Model has approximately 428B total parameters with 23B activated per token, using mixture-of-experts architecture.
• Supports text, image, and video inputs with a context window of up to 1,000,000 tokens.
• Uses MiniMax Sparse Attention throughout, plus quantization-aware training and a DSpark draft head for speculative decoding.
• Reports 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 82.2 on CyberGym.

Why it matters

Open weights for a large multimodal agentic model with long context and efficiency features could let developers run and study agentic coding and computer-use systems directly. The reported benchmarks are vendor-published comparisons against other models.

Read with care

Benchmark figures are self-reported by the model publisher and source claims not independently checked.