MiniMax M3 Powers AI Agents on NVIDIA

MiniMax M3 is a 428B parameter mixture-of-experts model capable of long-context reasoning and agentic workflows, featuring native multimodal input for video, image, and text across a 1M token context window, according to NVIDIA. The model delivers significantly improved performance on NVIDIA Blackwell Ultra infrastructure, as announced on June 12, 2026.

Enterprise AI developers often face challenges integrating disparate models for various data types. MiniMax M3 aims to simplify this by providing a unified multimodal system.

This single model facilitates applications such as long video understanding and extended coding sessions. It provides production-ready deployment paths on NVIDIA platforms.

MiniMax M3: A Multimodal Powerhouse

MiniMax M3 leverages a core architectural innovation called MiniMax Sparse Attention (MSA). This mechanism replaces standard quadratic attention by pre-filtering context blocks, only attending to the relevant ones NVIDIA reports.

Each Key-Value (KV) cache block is read once with contiguous memory access. This is over four times faster than existing sparse attention implementations. This design achieves 1/20th the per-token compute of its predecessor, M2, at a 1M-token context.

It also offers nine times faster prefill and 15 times faster decoding without compressing key-values or sacrificing precision. The model trains natively on text, images, and video from the start, using approximately 100 trillion interleaved tokens.

This approach avoids adding multimodality after initial training.

How Does Blackwell Ultra Accelerate Performance?

The NVIDIA Blackwell B300, specifically, delivers this increased throughput on MiniMax M3. This performance boost doubles interactivity and increases AI factory throughput, supporting more responsive inference.

The gains stem from co-design efforts integrating hardware and software components. These include speculative decoding, Multi-Tenant Processing (MTP), and accuracy-preserving NVFP4 acceleration the company explains.

Blackwell aims to further optimize MiniMax M3 performance as the software stack evolves. This particularly applies with NVIDIA Dynamo and NVIDIA CUDA kernels.

Developers can deploy MiniMax M3 using various open-source inference engines. These include NVIDIA TensorRT LLM, SGLang, and vLLM. NVIDIA provides configuration recipes for these engines.

For instance, deploying with TensorRT LLM involves using optimizations available on its GitHub repository. This setup includes downloading model checkpoints from Hugging Face and using a Docker container.

NVIDIA Dynamo is an open-source distributed inference serving platform. It helps deploy models like MiniMax M3 for large-scale applications. Dynamo integrates with major frameworks such as PyTorch, SGLang, TensorRT LLM, and vLLM.

It offers LLM-aware routing, elastic autoscaling, and low-latency data transfer. This integration allows for improved performance for long input sequence lengths without sacrificing throughput. For more on AI agents, see Microsoft's guide to building AI agents.

MiniMax M3 also allows customization and fine-tuning using the open-source NVIDIA NeMo Framework. Developers can use NVIDIA NeMo AutoModel for fine-tuning via SFT and LoRA over Hugging Face checkpoints. This process requires no conversion and benefits from high-throughput acceleration via N-D parallelism.

Context parallel support is available for sequence lengths up to 128k. Additionally, NVIDIA NeMo RL enables reinforcement learning on top of Minimax M3. These libraries provide tools for rapid experimentation on frontier models.