← Back to feed

MiniMax published the weights of MiniMax-M3.1-Flash to Hugging Face on October 23, completing the open-weight half of a rollout that began four days earlier when the model appeared on the company’s API as the former “Space Bunny” internal prototype. The repository, MiniMaxAI/MiniMax-M3.1-Flash, went live with the model card, license text, and day-one serving recipes for SGLang and vLLM.

Architecture and Serving Numbers

M3.1 Flash is a native multimodal mixture-of-experts model with approximately 428 billion total parameters and 23 billion activated per token. It accepts text, image, and video inputs with a one-million-token context window. Reasoning effort is always on, with levels from low to max.

Three structural changes separate it from June’s M3. First, MiniMax Sparse Attention now runs throughout the model, replacing the first three full-attention layers. Second, quantization-aware training is applied to the routed experts and attention paths during training — so the released checkpoint stores dense layers in MXFP8 and routed experts in NVFP4, rather than shipping raw weights for users to quantize later. Third, DSpark speculative decoding ships with a draft head inside the checkpoint itself, eliminating the need for a separate draft model at serve time.

The practical outcome is inference throughput exceeding 1,000 tokens per second on NVIDIA Blackwell and AMD Instinct GPUs through the SGLang Omni framework. MiniMax’s own service reports output speeds above 200 tokens per second for the hosted API, with a median above 170 — the Hugging Face release makes those same weights self-hostable.

Hardware support at launch:

• NVIDIA Blackwell (GB300): optimised through CUDA 13.4 QMUL4 instruction and NVFP4 MegaMoE fusion
• NVIDIA Rubin (SM107): dedicated Rubin kernels in vLLM
• AMD MI350 and MI355X: NVFP4-native MSA kernel and dual-batch overlap deliver 2-3x throughput gain over naïve serving
• Multi-chip via FlagOS for NVIDIA, AMD, and Chinese domestic silicon

Minimum viable deployment for full context and features requires four Blackwell or MI355X GPUs.

The License

The weights ship under the MiniMax Community License. Key terms:

• Non-commercial use: free, no conditions
• Commercial use below $20M annual revenue: permitted with a “Built with MiniMax M3.1-Flash” credit on the product’s website, interface, or documentation, plus a one-time notice to MiniMax
• Commercial use above $20M annual revenue: requires prior written authorization from MiniMax
• Military use and uses aimed at exploiting minors or spreading harmful disinformation: prohibited outright

The $20M revenue threshold matches the commercial gate in Mistral’s Modified MIT license and closely tracks the model-specific terms that have become a recurring pattern among Chinese frontier lab releases. The terms make M3.1 Flash freely accessible for startups and research but turn it into a negotiated agreement for enterprises above the threshold.

Pricing and Benchmarks

Hosted API pricing: $0.15 per million input tokens, $0.60 per million output tokens, $0.015 per million cached input tokens. No long-context surcharge applies above 512K tokens, unlike MiniMax-M3’s standard pricing structure.

Vendor-reported benchmark scores from the launch materials:

Benchmark	Score
CyberGym	82.2
BankerToolBench	76.2
Inference throughput (Blackwell, SGLang)	>1,000 tok/s

Independent verification of the task-level benchmark figures has not yet been published.

Serving Stack at Launch

vLLM and SGLang ship with day-one support. Docker Model Runner and TokenSpeed are also listed. Transformers, KTransformers, Unsloth, and ATOM integrations are in progress. The basic serving path:

huggingface-cli download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash
pip install vllm
vllm serve "MiniMaxAI/MiniMax-M3.1-Flash"

Context

M3.1 Flash had processed over 200 trillion tokens as Space Bunny before the API launch on October 19. The open-weight release is the first time the underlying checkpoint is available for inspection, fine-tuning, or self-hosted deployment. MiniMax’s gross margin for the first half of 2026 was 17.9%, a figure that constrains how aggressively the company can undercut on API pricing — the open-weight release shifts some of that pressure to users with their own infrastructure.