Together AI launches MiniMax M3 with 1M-token context and native multimodality

Together AI has announced a partnership with MiniMax and put the M3 model into production, with a context window of up to one million tokens, native multimodality and up to 125% faster inference.

What is MiniMax M3

MiniMax M3 is the company's flagship model, built for real agentic tasks: long documents, codebases, images and tool calls, all in a single context. Earlier models struggled with long contexts because of the quadratic complexity of self-attention. M3 takes a fundamentally different approach.

The architecture is built on MiniMax Sparse Attention (MSA), a block-sparse attention mechanism. Each query token attends only to a limited number of relevant KV-cache blocks rather than the whole sequence. As a result, computational cost no longer grows quadratically with context length. Compared with M2.7, prefill is 9 times faster and decoding is 15 times faster.

Key characteristics of M3:

• Context window of up to 1M tokens
• Native multimodality: text, code, images
• Support for agentic scenarios and tool calls
• Competitive results in programming tasks

How Together AI prepared the infrastructure

Running a model like this in production without losing efficiency is a non-trivial task. MSA computes attention in two stages: first it calculates relevance to select KV blocks, then it applies dense attention between the query tokens and the selected blocks. Standard implementations cannot handle this well.

The Together AI engineering team wrote four key optimizations:

• KV-Block-Major sparse attention kernel: a CUDA kernel for block-wise sparse attention over the KV cache, with memory data layout reorganized for MSA's specifics
• Paged MSA decode: integration of paged attention for decoding at million-token context without memory fragmentation
• Optimized index scoring kernel: faster relevance calculation for selecting KV blocks at every decoding step
• Rust-based multimodal gateway: a preprocessing gateway for images and mixed inputs with minimal latency

Together, these optimizations raised throughput by 81% to 125% at different load levels. Measurements were taken on NVIDIA B200.

What comes next

MiniMax M3 is already available through Together AI as a cloud service. Open weights will be released in the coming days. After that, the model can be deployed independently or used as an API endpoint directly on the Together platform.

"The production launch of M3 confirms

Together AI as the preferred platform for models that pose real systems challenges," from the company's official blog.

What this means

Together AI's ability to run a model with a 1M-token context 81–125% more efficiently than the standard approach is a competitive signal to the market. For developers, it means near-term access to a powerful multimodal agent through a simple API, without having to work through the architectural details of MSA themselves.