Together AI launches MiniMax M3 with 1 million-token context and multimodal support

Together AI Launched MiniMax M3 with 1 Million Token Context and Multimodality

Together AI announced a partnership with MiniMax and launched the M3 model in production — with context up to one million tokens, native multimodality, and inference acceleration up to 125%.

What is MiniMax M3

MiniMax M3 is the company's flagship model, created for real-world agentic tasks: long documents, codebases, images, tool invocations — all in one context. Previous models struggled with long contexts due to quadratic complexity of self-attention. M3 solves this problem in a fundamentally different way.

At the core of the architecture is MiniMax Sparse Attention (MSA), a block-sparse attention mechanism. Each query token attends to only a limited number of relevant KV-cache blocks rather than the entire sequence. Because of this, computational complexity no longer grows quadratically with context length.

Compared to M2.7: prefill acceleration is 9x, decoding acceleration is 15x. Key M3 characteristics:

• Context window — up to 1 million tokens
• Native multimodality: text, code, images
• Support for agentic scenarios and tool invocations
• Competitive results on programming tasks

How Together AI Prepared the Infrastructure

Launching such a model in production without losing efficiency is non-trivial. MSA assumes two stages when computing attention: first, calculating relevance to select KV-blocks, then dense attention between query tokens and selected blocks. Standard implementations do not handle this well.

The Together AI engineering team wrote four key optimizations:

• KV-Block-Major sparse attention kernel — a CUDA kernel for sparse attention over KV-cache blocks with memory reorganization specific to MSA
• Paged MSA decode — integration of paged attention for decoding with million-token context without memory fragmentation
• Optimized index scoring kernel — accelerated relevance calculation for KV-block selection at each decoding step
• Rust-based multimodal gateway — a preprocessing gateway for images and mixed inputs with minimal latency

In total, these optimizations delivered throughput gains from 81% to 125% at various load levels — measurements were taken on NVIDIA B200.

What's Next

MiniMax M3 is already available through Together AI as a cloud service. In the coming days, open weights will be released — after which the model can be deployed independently or used as an API endpoint on the Together platform directly.

"Launching M3 in production confirms

Together AI as the preferred platform for models that impose real system-level demands," — from the company's official blog.

What This Means

Together AI's ability to deploy a model with 1M-token context 81–125% more efficiently than a standard approach is a competitive signal to the market. For developers, it means quick access to a powerful multimodal agent via a simple API without needing to work through the architectural nuances of MSA yourself.