Models Published:
MiniMax publishes M3.1-Flash weights: a 428B multimodal MoE with a 1M-token context and a 252 GB download
What happened
MiniMax released the open weights of MiniMax-M3.1-Flash on Hugging Face on October 23. It is a native multimodal mixture-of-experts model with about 428B total parameters and 23B active per token, taking text, image and video input with a context window of up to 1M tokens, and it is aimed at coding, agent workflows, computer use and cybersecurity. Compared with MiniMax-M3 it uses sparse attention in every layer, quantization-aware training and a new DSpark speculative-decoding draft head, which ships with the weights. MiniMax recommends SGLang, vLLM or TokenSpeed to serve it, and the weights come under the MiniMax Community License rather than an open-source license.
Why it matters locally: It is another open-weight agent model in the same class as DeepSeek V4.1 Flash, but its license asks commercial users to show a "Built with MiniMax M3.1-Flash" credit and send MiniMax a one-time notice, or get prior written approval above US$20 million in yearly revenue.
Three ways to read it
For the market
MiniMax's own table puts M3.1-Flash level with Opus 5 on BankerToolBench (76.2) but behind DeepSeek V4.1 Flash-1017 on Terminal Bench 4.0 (38.1 versus 42.3), so the two Chinese labs look closely matched on agent work. The deciding factor for many companies will likely be the license: DeepSeek's MIT terms are simpler than MiniMax's notice-and-approval rules.
In plain words
MiniMax, a Chinese AI company, has made one of its big AI models free to download, and it can read text, pictures and videos and work through long multi-step jobs. The files add up to about 252 GB, so it needs a very large computer, and businesses that use it to make money have to tell MiniMax first.
For practitioners
The Hugging Face repo totals about 252 GB, so plan for more than 256 GB of memory once the KV cache for long contexts is added: an M5 Ultra Mac Studio in the 512 GB configuration Apple has announced, or a multi-GPU server, is the realistic starting point. Use SGLang, vLLM or TokenSpeed as MiniMax recommends (Transformers, KTransformers and unsloth support are still in progress), start from temperature 1.0 and top_p 0.95, and test whether the DSpark draft head speeds up your own workload.
Sprout's insight
Sprout compares this with other brands and with where real applications are heading. Free for members.
Sign in free to read it →Subscribe to Sprout's weekly
One email a week: the top stories and a taste of Sprout’s insights. Unsubscribe any time.
Source
Read the official source: MiniMax-M3.1-Flash model card (Hugging Face) ↗Summaries are written by OnPrem World from the official source. We link to the original and never republish it.
On this site
More news
- NVIDIA's small Nemotron 3.5 Bolt E4B arrives in Ollama 0.40.10, with a DFlash draft for faster replies
- Ollama 0.40.9 adds local text-to-speech on Apple Silicon, d3 decision models, and Flint on Linux and Windows
- Recursive releases Flint, an Apache 2.0 coding model with 40B parameters and about 2.5B active, already in Ollama for Mac
- AMD and Intel agree on oneDNN as PyTorch's shared x86 CPU acceleration library
- vLLM v0.32 brings 4-bit KV cache to more GPUs and stops allowing every CORS origin by default