MiniMax Releases M3: 428B-Parameter Multimodal Model with 1M Context Window and 15× Decode Speedup

MiniMax has released M3, a multimodal model with approximately 428 billion parameters and 23 billion activated parameters. The model supports a 1 million token context window and uses MiniMax Sparse Attention (MSA) to achieve 9× prefill and 15× decode speedups compared to its predecessor M2, reducing per-token compute to 1/20.

Technical Specifications

M3 uses native multimodal training from the first step, processing text, image, and video inputs through mixed-modality training rather than adapting a text-only model. The model employs MiniMax Sparse Attention, which the company claims dramatically reduces attention compute and memory footprint compared to Grouped Query Attention (GQA) while preserving model quality.

The model features two operating modes: a "thinking" mode for complex reasoning and agentic tasks, and a "non-thinking" mode for latency-sensitive scenarios like chat and code completion. According to MiniMax, M3 achieves frontier-level performance across long-horizon agentic benchmarks.

Pricing details have not been disclosed. The model is available through the MiniMax API and for local deployment via Hugging Face.

Deployment Options

M3 can be deployed locally using three inference frameworks: SGLang, vLLM, and Transformers. MiniMax recommends specific inference parameters: temperature=1.0, top_p=0.95, and top_k=40.

The model supports API access through MiniMax's own API service, with Novita listed as an additional inference provider on Hugging Face. The technical details are available in a research paper on arXiv (arXiv:2606.13392).

What This Means

M3's sparse attention architecture addresses a critical bottleneck in long-context models: compute cost at scale. The claimed 15× decode speedup at 1M tokens, if validated in independent benchmarks, would make M3 significantly more practical for production use cases requiring extended context.

The native multimodal training approach contrasts with common industry practice of adapting text models for visual inputs. This architectural choice suggests MiniMax is betting on deeper semantic integration across modalities, though real-world performance comparisons with models like GPT-4o or Gemini 1.5 Pro will determine whether this approach delivers measurable advantages. The emphasis on agentic capabilities and coding performance positions M3 as a competitor in the autonomous agent and development tools market.