MiniMax published the M3.1 Flash checkpoint on October 23, 2026, creating a self-hosting path after its earlier API introduction. The release is visible in the official repository’s dated initial commit, which includes sharded model weights and the DSpark draft component. This milestone concerns downloadable artifacts, rather than a new API launch.
What the weight release contains
The official model card describes a native multimodal mixture-of-experts model with approximately 428 billion total parameters and 23 billion activated per token. It accepts text, images and video and produces text, with a stated context ceiling of one million tokens.
MiniMax describes sparse attention throughout, quantization-aware training and an included DSpark speculative-decoding head. Its deployment section recommends SGLang, vLLM and TokenSpeed, while listing other integrations as work in progress. Automatically generated library examples should not be treated as proof that every framework supports the complete checkpoint.
The repository history dates the artifact-bearing initial commit to October 23. The repository’s earlier creation timestamp is a different event and does not establish that weights were public then.
Open weights carry specific commercial conditions
The checkpoint uses the MiniMax Community License, rather than Apache 2.0. The license sets a commercial attribution requirement and a one-time notice requirement below its revenue threshold. Products and services generating more than US$20 million in yearly revenue require separate prior written authorization.
The text also includes prohibited uses and preservation of notices. A business should check the full terms against its intended deployment before treating a public download as unrestricted commercial permission. Hosted API purchasing and weight licensing are separate decisions.
Active parameters do not describe the serving footprint
The vLLM recipe documents MXFP8 dense layers and NVFP4 experts, with hardware-specific serving paths. Its examples include four-way tensor parallelism on Blackwell or AMD MI355X systems and eight-way parallelism on Hopper. These are substantial accelerator configurations, not a laptop installation guide.
Capabilities also vary by recipe. The documented Hopper example limits context to 262,144 tokens, disables video input and does not enable DSpark. The model’s maximum advertised context and modalities therefore do not establish what a particular deployed endpoint accepts.
What this changes for an engineering team
Our assessment is that downloadable weights create greater control over hosting and version selection, while transferring more responsibility for operations to the deployer. A team must own artifact verification, runtime compatibility, access control, updates and capacity. Public weights do not perform those tasks automatically.
A useful first evaluation would compare the intended hosted route and a pinned self-hosted revision on the same bounded maintenance tasks. Record accepted changes, reviewer time, memory demand and elapsed time. Include failed runs in the comparison so retries remain visible in the operating cost. Keep tool permissions constant so that changes in the execution environment do not masquerade as model improvements.
MiniMax’s reported benchmark results should remain attributed to MiniMax; they are not Nerova measurements and do not establish a particular organization’s cost per accepted change. The October 23 release makes a concrete deployment evaluation possible, with license conditions and hardware configuration included in that evaluation from the start.