# News: MiniMax publishes M3.1 Flash weights: 428B-parameter multimodal MoE for coding and agents, 1M-token context

> MiniMax published open weights for M3.1 Flash on 23 October 2026, four days after the model's hosted launch on 19 October. The Hugging Face model card confirms a native multimodal mixture-of-experts model: approximately 428 billion total parameters with 23 billion activated per token, a 1,000,000-token context window, text, image and video inputs, sparse attention throughout, and a DSpark speculative-decoding draft head, under the MiniMax Community licence (not a permissive Apache-style licence). MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench and 82.2 on CyberGym, all vendor-run, and says the hosted model costs 70% less than its predecessor with 200+ output tokens per second. Weights are live on Hugging Face and ModelScope, community GGUF conversions appeared the same day, and MiniMax published deployment guidance for SGLang and vLLM.

Human page: https://testedactually.com/news/minimax-m31-flash-open-weights/
Mirror: https://testedactually.com/news/minimax-m31-flash-open-weights.md
Published: 2026-10-25

## What changed

MiniMax published open weights for M3.1 Flash on 23 October 2026, four days after the model's hosted launch on 19 October. The Hugging Face model card confirms a native multimodal mixture-of-experts model: approximately 428 billion total parameters with 23 billion activated per token, a 1,000,000-token context window, text, image and video inputs, sparse attention throughout, and a DSpark speculative-decoding draft head, under the MiniMax Community licence (not a permissive Apache-style licence). MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench and 82.2 on CyberGym, all vendor-run, and says the hosted model costs 70% less than its predecessor with 200+ output tokens per second. Weights are live on Hugging Face and ModelScope, community GGUF conversions appeared the same day, and MiniMax published deployment guidance for SGLang and vLLM.

## What it means for buyers

At 428B total parameters this is a multi-GPU server or high-memory workstation model, not a laptop download, so Ollama and LM Studio users should size hardware before benchmarking - the release matters most for self-hosting teams on SGLang or vLLM who now have a low-cost agentic and coding model outside the US labs, and for buyers weighing self-hosted inference against MiniMax's hosted API. The benchmarks above are vendor-reported; validate on your own tasks before switching. It lands in the same open-weight wave as Kimi K3.1 Pro, which targets the same coding-and-agents workload at a smaller footprint.

## Sources

- [1] MiniMax-M3.1-Flash model card (Hugging Face) - https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash (weights release, parameters, context window, licence, benchmarks, checked 2026-10-25)
- [2] MiniMax M3.1 Flash: Redefining Agent Performance and Cost (MiniMax blog) - https://www.minimax.io/blog/minimax-m31f (hosted launch 19 October, cost claim, availability, checked 2026-10-25)
- [3] MiniMax releases open weights for M3.1 Flash (iMasters) - https://imasters.com/news/minimax-releases-open-weights-for-m31-flash-a-428-billion-parameter-model (independent report dating the weight drop 23 October, SGLang and vLLM guidance, checked 2026-10-25)

TestedActually tracks pricing and programme changes as dated, sourced events: https://testedactually.com/news/.