The stamp at the top of this IT Home page is 2026/10/23 9:08:08. MiniMax said today, in the page's wording, that it is open-sourcing the weights of MiniMax M3.1 Flash. Total parameters are 428 billion. Active parameters are 23 billion. Those two counts are the product specification printed on this page. The page says that, compared with the prior M3, M3.1 Flash is fully improved. Fully improved is the company's wording. This page prints no comparison scores, so the sentence cannot be turned into a scoreboard.

[1]
A charcoal weight, blank top, muted blue handle
A charcoal weight with a blank top, a muted blue handle, and a small oxide-red mark on the right., AI-generated illustration, not a news photograph

The launch date is separate. M3.1 Flash went live on October 19, and within 72 hours real users processed more than 100 trillion tokens. The 100 trillion is a usage figure on the page, not a capability score. It does not fill in the comparison table that the fully-improved sentence lacks. The open-weight announcement is today on this page. The go-live date is October 19. They are not the same day.

[1]

Day-one access is a list, not a setup procedure. MiniMax says it is working with Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, MetaX, Tianshu, NVIDIA, and AMD, and with Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean, on day-one open access. vLLM, SGLang, FlagOS, Hugging Face, and ModelScope are also named, to place the model on different chips, clouds, and development setups. The page gives the Hugging Face path MiniMaxAI/MiniMax-M3.1-Flash and the ModelScope path MiniMax/MiniMax-M3.1-Flash. The list and the paths only record the cooperation the page claims.

[1]

The page also mentions sparse attention, quantization-aware training, and power scheduling. Those sentences describe how the structure was changed, how the numbers were moved, and how power was assigned. This story does not repeat those steps and does not turn them into a configuration. What remains is the judgment: 428 billion and 23 billion are the specification, 100 trillion tokens are usage in the 72 hours after launch, and the claim of a full improvement has no scores. Scoreboards and deployment notes for other open models at the bottom are other stories and are not folded into M3.1 Flash.

[1]