MiniMax has released the weights for M3.1 Flash, a multimodal Mixture-of-Experts model with 428 billion total parameters, of which 23 billion are active per token. The model supports text, images, and video, alongside a context window of up to 1 million tokens. It is engineered specifically for code generation, autonomous agents, computer use, and cybersecurity. Key architectural modifications include sparse attention across all layers, quantization-aware training, and a DSpark attention head designed for speculative decoding relative to the previous M3 iteration.
According to MiniMax’s internal benchmarking, the model scores 73.7% on DeepSWE v1.1 compared to a competitor score of 74.0%. On OSWorld 2.0, it registers 61.7% against 70.6% for the competitor, while Terminal Bench 4.0 yields a substantially lower result of 38.1% versus 51.8%. Infrastructure deployment can be executed locally via SGLang or vLLM using out-of-the-box recipes. This capability directly pressures cloud providers whose business models rely heavily on API inference monetization. When development teams gain the capacity to deploy models of this scale on proprietary hardware, the unit economics of inference shift fundamentally. Bypassing hourly or per-token proprietary API tariffs significantly cuts operational expenditure on long-context processing and autonomous agent workflows, reallocating budgets from cloud rental to internal infrastructure.
Key points
Multimodal MoE architecture featuring 428B total and 23B active parameters with a 1M token context window. Native support for local deployment via SGLang and vLLM, cutting reliance on hosted APIs. Proprietary MiniMax Community license creates legal compliance hurdles for commercial pipelines.
Editorial analysis
How the story developed
On 1 June 2026, MiniMax introduced the multimodal M3 with a 1M-token context window, scoring 59% on SWE-bench Pro ahead of GPT-4o, and on 20 August released open weights and checkpoints for its Ling-3.0-flash and Ling-3.0-tiny lines across public platforms. Now the company has published open weights for the 428B-parameter MoE model M3.1 Flash with 23B active parameters per token, upgrading the architecture with full-layer sparse attention, quantization-aware training, and a DSpark speculative decoding head. MiniMax is systematically shifting its open-weight strategy from raw multimodal scale toward execution-focused agent workflows and inference efficiency.
Sources: our stories of 01.06.2026, 20.08.2026