最新模型
MiniMax Releases M3.1-Flash Weights, Integrating All-Layer Sparse Attention and DSpark Speculative Decoding
MiniMax released the M3.1-Flash weights on October 23, featuring all-layer sparse attention, quantization-aware training, and a built-in draft head. Deployment docs list configuration differences across GPUs; the million-token context and speculative decoding still require hardware validation.
On October 23, MiniMax created the official M3.1-Flash weights repository on Hugging Face, making the multimodal model available for self-hosted deployment. The model has about 428 billion total parameters, with roughly 23 billion activated per token. It accepts text, images, and video as input and outputs text; the company lists a maximum context of up to one million tokens. Weight release history, model card
Compared with M3, the new version replaces full attention in the first three layers with sparse attention, extending MSA across all layers. The routed experts and attention path use quantization-aware training, and DSpark replaces the previous EAGLE-like multi-token prediction approach. The release includes a draft head, but whether it actually speeds up generation depends on serving framework and hardware support. Technical changes
The vLLM deployment recipe shows that the weights use MXFP8 for dense layers and NVFP4 for experts. Blackwell can run NVFP4 experts natively, while H100/H200 use Marlin kernels. The Hopper example uses eight GPUs, an FP8 cache, and a 262,144-token limit; it disables video input and does not enable DSpark. These differences mean engineering teams cannot treat the model card’s capability ceiling as the service specification their current cluster can deliver. vLLM deployment recipe
Agent integration also has a specific requirement: SGLang documentation specifies the minimax-m3 tool parser to obtain structured tool_calls. Reasoning mode can be enabled, disabled, or set to adaptive. Deployment acceptance testing should cover the separation of tool arguments, reasoning content, and answers, and measure latency and memory use with long contexts. SGLang documentation
The license also affects adoption. The community license requires commercial products to identify the model by name. Products and services with annual revenue exceeding $20 million require prior written authorization; other commercial use requires sending a one-time notification. Key areas to watch include independent replication across hardware configurations and whether third-party quantization preserves the task performance of the official weights. License terms