Loading page…
The public download brings self-hosting to a model previously offered as a paid service. Its smaller files still require substantial hardware, and commercial deployments face attribution and permission rules.
Listen to this story
MiniMax has made M3.1 Flash’s roughly 252 GB of weights available on Hugging Face, creating a self-hosting option for teams able to provision multi-GPU servers. The model has 428 billion total parameters, though it activates 23 billion per token. Commercial deployment is conditional: products must attribute and notify MiniMax, while organizations above $20 million in annual revenue need prior written authorization. The release enables independent testing, but published benchmark scores remain unverified, and hosted API prices do not establish self-hosting costs.
MiniMax’s vLLM deployment recipes require eight Hopper GPUs, four Blackwell-class GPUs, or four AMD MI355X GPUs; DSpark is disabled on Hopper.
The model accepts text, images, and video, and supports a context window of up to one million tokens.
M3.1 Flash’s repository is about one-third the size of M3’s, although both models have the same total and active parameter counts.
MiniMax’s M3.1 Flash can now run outside the company’s servers—provided the operator has substantial hardware and accepts its license conditions. MiniMax published the model’s weights on Hugging Face on October 23, 2026, with no access request required. The roughly 252 GB download opens a self-hosted route, but not an unrestricted one.
The release follows M3.1 Flash’s October 19 debut as a paid service. MiniMax had promised the weights for October 23. Publishing them supplies the trained numerical values needed to run the model independently, rather than sending requests through the company’s hosted interface.
MiniMax’s model card lists about 428 billion parameters, with 23 billion active for each token, or unit of processed text. Its mixture-of-experts design selects specialized groups of parameters instead of using the entire model for every token. The smaller active count describes the work performed, not the full model that must be stored.
The model accepts text, images and video, and produces text. Its context window—the material it can work with in one request—holds up to one million tokens. That input capacity is separate from the parameter count: one describes the request, while the other describes the trained model.
MiniMax uses quantization-aware training, which incorporates reduced numerical precision during training. The release also includes DSpark speculative decoding: a lightweight component proposes several tokens, and the larger model checks them. That approach is intended to speed text generation. Together with sparse attention, these are central changes from the earlier M3 model.
The official deployment recipe is not a laptop setup. For vLLM, software used to serve the model, MiniMax specifies one of three multi-GPU configurations. It also requires version 0.33.0 or later and a block size of 128 for sparse attention. DSpark is not enabled in the Hopper configuration.
Hosted access remains an alternative for testing without that equipment. MiniMax lists rates of $0.15 per million input tokens and $0.60 per million output tokens. Those are service prices, not estimates of what operating the downloadable model will cost.
The MiniMax Community License permits free non-commercial use, but commercial use requires attribution and notification to MiniMax. Above $20 million in annual revenue, prior written authorization is required. Publicly downloadable weights therefore do not mean unrestricted commercial use.
Commercial products must prominently display the attribution “Built with MiniMax M3.1-Flash.” The license’s commercial definition includes paid services built on the model, commercially served APIs, and commercial distribution of modified or further-trained versions. Below the revenue threshold, the notification is a one-time email to MiniMax.
Performance is a separate consideration from access. MiniMax’s published results include 73.7 on DeepSWE 1.1 and 38.1 on Terminal Bench 4.0. Hugging Face marks those evaluation entries as unverified. The weight release gives developers a way to test the model themselves; it does not independently validate the company’s benchmark results.
Loading discussion...
Join the conversation
Explain whether downloadable access outweighs the licensing uncertainty.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.