MiniMax opened the weights of M3.1 Flash on October 23, 2026. In an official announcement, MiniMax said the M3.1 Flash weights are now public: 428B total parameters with about 23B activated per token, a mixture-of-experts (MoE) architecture, and a 1-million-token context window. The model launched as an API on October 19 and had previously been tested anonymously under the codename "Space Bunny"; with the weights out, developers can self-host instead of going through MiniMax's API.
Three bets on inference efficiency
The technical changes MiniMax describes center on three things: full-model sparse attention, which cuts the cost of attention over long contexts; quantization-aware training (QAT), which keeps accuracy loss smaller at low precision; and power-aware scheduling, which matches compute to load. Combined with sparse activation of only 23B parameters per token, the model's API pricing is $0.15 per million input tokens, a flat 24/7 rate with no peak/off-peak split.
Chinese chips take up half the day-one list
The deployment ecosystem arrived on day one: on silicon, Huawei Ascend, T-head, Kunlunxin, Moore Threads, Biren, Hygon, Musen and TianShu, alongside NVIDIA and AMD; on cloud, Together AI, Fireworks, Nebius, Baseten, Modal and DigitalOcean; on tooling, vLLM, SGLang, Baidu's FlagOS, Hugging Face and ModelScope. Being installable is treated as seriously as scoring well.
Why this matters
Open weights for a 428B-scale MoE are rare enough; the cadence is the real signal. Four days from API launch to open weights suggests open models are shifting from "open at release" to "validate demand first, release weights later" — the Space Bunny stealth period processed over 200 trillion tokens (MiniMax's figure), meaning the model was battle-tested on real traffic before being handed to the community. For developers in China, day-one support across domestic chips means no waiting for community ports to run it locally. The caveats: a 428B backbone, sparse or not, is still a multi-GPU workstation or server job, not consumer hardware; and with no technical report published, what changed from M3 to Flash remains MiniMax's own account.
