Ted Hisokawa
Jun 12, 2026 15:13
MiniMax M3, the 428B-parameter model, launches on NVIDIA infrastructure, offering long-context reasoning and multimodal workflows for enterprise AI.
MiniMax M3, a cutting-edge 428-billion-parameter AI model, is now available on NVIDIA’s accelerated infrastructure, including its Blackwell GPUs. The model, released by Shanghai-based MiniMax on June 1, 2026, aims to simplify enterprise AI workflows by combining long-context reasoning, multimodal capabilities, and agentic task optimization—all in a single system.
The standout feature of MiniMax M3 is its ability to process up to 1 million tokens in context, a massive upgrade over most existing models. This enables extended coding sessions, complex legal document analysis, or long-form video understanding without breaking context. Additionally, the model supports native multimodal input—text, images, and video—eliminating the need for separate pipelines and reducing complexity for developers.
Architectural Advances: MiniMax Sparse Attention
At the heart of M3’s performance is the new MiniMax Sparse Attention (MSA) architecture. Unlike traditional quadratic attention mechanisms, MSA uses a pre-filtering stage to focus only on relevant context blocks, dramatically improving speed and efficiency. According to MiniMax, this reduces computational costs to just 1/20th of its predecessor, MiniMax M2, for 1M-token contexts. Prefill speeds are reportedly nine times faster, while decoding is 15 times faster compared to older sparse attention implementations.
The model also trains natively across text, images, and video from the ground up, with no need for post-training multimodality hacks—a key differentiator in the frontier model space.
Enterprise Deployment and Customization
The MiniMax M3 can be deployed using popular open-source inference engines like NVIDIA TensorRT LLM, SGLang, and vLLM. NVIDIA has integrated the model into its Dynamo distributed inference platform, which enhances performance for long-sequence workloads by separating prefill and decode tasks across GPUs. This approach reportedly delivers a 4x improvement in interactivity at 32k input length sequences on NVIDIA Blackwell hardware.
For those looking to customize M3, NVIDIA’s NeMo Framework offers robust tools for fine-tuning, including support for sequence lengths up to 128k tokens. Developers can also perform reinforcement learning with the model to optimize it for specific applications like agent-based workflows or document parsing.
Competitive Market Position
MiniMax M3 is entering a crowded AI model market but aims to differentiate itself through its technical capabilities and open-weight approach. On coding benchmarks, MiniMax claims a 59.0% score on SWE-Bench Pro, narrowly outperforming GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). While these results are company-reported, they position M3 as a leading contender in the coding and multimodal AI space.
Crucially, the model undercuts many closed-source competitors on cost, with pricing reported at $0.60 per million input tokens at launch. This aggressive pricing strategy targets cost-sensitive enterprises deploying large-scale AI workflows.
What’s Next?
Developers can start working with MiniMax M3 immediately via NVIDIA’s GPU-accelerated API or by downloading model weights from Hugging Face. With its open-weight design, the model is expected to see wide adoption in domains like legal tech, autonomous systems, and multimodal content generation.
While the AI world will be watching closely to verify MiniMax’s claims on efficiency and benchmarks, the model’s technical innovations and cost structure make it a compelling option for enterprises looking to streamline complex workflows.
Image source: Shutterstock
Be the first to comment