As enterprise AI adoption scales, builders are more and more compelled to sew collectively fragmented pipelines—separate fashions for textual content, imaginative and prescient, and code—resulting in added complexity, greater prices, and slower iteration.
MiniMax M3—accessible on NVIDIA accelerated infrastructure together with NVIDIA Blackwell—adjustments this by enabling a single multimodal system able to long-context reasoning, agentic workflows, and artistic duties.
The 428B parameter MoE helps as much as 1M tokens and native multimodal enter. Builders can construct functions like lengthy video understanding, prolonged coding classes (8+ hours), and high-quality design workflows—all with a unified mannequin and production-ready deployment paths on NVIDIA platforms.
MiniMax M3’s core architectural innovation is MiniMax Sparse Consideration (MSA), which replaces normal quadratic consideration with a pre-filtering stage that identifies related context blocks and attends solely to these. On the operator degree, every KV cache block is learn as soon as with contiguous reminiscence entry—greater than 4x sooner than present sparse consideration implementations. This yields 1/twentieth the per-token compute of M2 at 1M-token context, with 9x sooner prefill and 15x sooner decoding, all with out compressing key-values or sacrificing precision. The mannequin additionally trains textual content, pictures, and video natively from step 0 throughout ~100 trillion interleaved tokens, reasonably than including multimodality post-training.
Open supply inference
Builders can use accelerated computing with their open supply inference engine of alternative, corresponding to NVIDIA TensorRT LLM (text-only), SGLang or vLLM.
Deploying with NVIDIA TensorRT LLM
The optimizations can be found on the NVIDIA TensorRT LLM GitHub repository. Observe the fast begin information to face up a high-performance server—it covers downloading mannequin checkpoints from Hugging Face, a ready-to-run Docker container, and configuration choices for each low-latency and max-throughput serving. NVIDIA additionally collaborated on the developer expertise by the Transformers library.
Deploying with SGLang
Customers deploying fashions with the SGLang serving framework can use the next directions. See the SGLang documentation for extra info and configuration choices.
$ python -m sglang.launch_server
–model-path MiniMaxAI/MiniMax-M3
–dtype bfloat16
–tp-size 8
–ep-size 8
–trust-remote-code
–mem-fraction-static 0.8
–enable-multimodal
–quantization mxfp8
–attention-backend flashinfer
–mm-attention-backend flashinfer_cudnn
–moe-runner-backend deep_gemm
–chunked-prefill-size 8192
–reasoning-parser minimax-m3
–tool-call-parser minimax-m3-nom
–tr
Deploying with vLLM
When deploying fashions with the vLLM serving framework, use the next directions. For extra info, see the vLLM Recipe.
–tensor-parallel-size 8
–enable-expert-parallel
–block-size 128
–mm-encoder-attn-backend FLASHINFER
–mm-processor-cache-type shm
–tool-call-parser minimax_m3
–enable-auto-tool-choice
–reasoning-parser minimax_m3
–trust-remote-code
Scaling with NVIDIA Dynamo
Dynamo is an open supply distributed inference serving platform for builders to deploy frontier fashions like MiniMax M3 for large-scale functions. Deploying MiniMax M3 utilizing Dynamo with TensorRT LLM improves efficiency for lengthy enter sequence lengths with out sacrificing throughput or rising GPU finances.
Dynamo integrates with all main inference engines and frameworks, together with PyTorch, SGLang, TensorRT LLM, and vLLM, and presents LLM-aware routing, elastic autoscaling, and low-latency information switch. Builders can comply with the deployment information to run MiniMax M3 with Dynamo.
Customise with NVIDIA NeMo Framework
MiniMax M3 could be custom-made and fine-tuned with the open supply NVIDIA NeMo Framework. Customers can:
Use NVIDIA NeMo AutoModel for out-of-the-box fine-tuning (each SFT and LoRA) over Hugging Face checkpoints with out any conversion, with high-throughput acceleration from full N-D parallelism. Particularly, context parallel assist is accessible for sequence lengths as much as 128k.
Use NVIDIA NeMo RL to conduct reinforcement studying on high of Minimax M3, referencing the next pattern accuracy curves.
These libraries present builders with a suite of light-weight instruments for speedy experimentation on the most recent frontier fashions.
Get began at present
Builders can prototype and consider MiniMax M3 by utilizing the GPU-accelerated API on construct.nvidia.com or by downloading the weights from Hugging Face.
