Every model that forced you to chunk a 500-page codebase into fragments before analysis just became harder to justify.
The chunking tax on long-context work is real and expensive
Engineers and researchers who work with long documents, massive codebases, or multi-hour video transcripts have had to split inputs, manage context windows manually, and stitch outputs back together. That workflow introduces errors, costs time, and breaks reasoning chains that depend on the full document being visible at once.
One model now reads a million tokens, sees images, and writes code
MiniMax Releases MiniMax M3 with MSA Architecture Supporting 1M accepts text, images, and video as native inputs, processes up to 1 million tokens in a single context window using its new MSA (MiniMax Sparse Attention) architecture, and outputs code, analysis, or desktop computer actions through the live API. You connect via the MiniMax API, MiniMax Code, or the MiniMax Token Plan, paste or pipe your input, and get a single coherent response across content types. Model weights and the full technical report are scheduled for release within 10 days of launch.
Three roles feel this most immediately
- Software engineers who need to audit an entire monorepo for security vulnerabilities without losing cross-file context between passes
- Research analysts who ingest multi-hundred-page regulatory documents and need citations that reference the correct section without hallucinated page numbers
- AI infrastructure teams evaluating open-weight frontier models who need multimodality, long context, and agentic capability in a single deployable package
The MSA architecture is the detail worth watching here. By partitioning the KV cache into precisely bounded blocks and using a KV-outer-gather-Q operator that reads each block once with contiguous memory access, MiniMax reports over 4x speed improvement versus existing open-source sparse attention implementations.
The open-weight frontier just moved past what closed APIs were offering six months ago
Gemini 1.5 Pro shipped a 1M-token window in 2024 as a closed API feature that analysts cited as a structural moat. MiniMax M3 releasing that same capability as open weights, bundled with native video input and desktop operation, compresses that moat significantly and gives self-hosting teams options they did not have before the June 2025 release window.
What you can actually do with this today
- Feed an entire codebase into one prompt and ask for a refactor plan
- Pass a full video file and extract structured event timelines natively
- Run agentic coding tasks that operate on your desktop without external tooling
- Deploy the open weights on your own infrastructure once released within 10 days
Pricing is usage-based through the MiniMax API and Token Plan — check the MiniMax site for current rate cards.
The honest ceiling
Weights are not yet public as of launch day, so independent benchmarking of the MSA architecture claims cannot happen until the scheduled release window closes.
The open-weight long-context race has real contenders now
Qwen2.5 covers long-context text well but does not natively handle video input or desktop operation. If multimodal agentic behavior in a single open-weight model is the requirement, MiniMax M3 currently has no direct open-weight competitor that matches all three capabilities simultaneously.
Open-weight multimodal models are closing the gap on closed API leaders faster than expected
This is the kind of architectural shift we track before it becomes consensus. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.