MiniMax M3: The Open-Weights Model Making 1M Context Affordable

On June 1, 2026, MiniMax (Xiyu Technology) released M3, and the pitch is unusually ambitious: the first open-weights model that combines frontier-level coding, a 1-million-token context window, and native multimodal input in a single system. What makes it different isn't just the specs—it's the price. At promotional rates of roughly $0.30 per million input tokens and $1.20 per million output tokens, M3 runs long-context agentic tasks at about 5% of what you'd pay Claude Opus for the same work.

For developers building AI agents, this changes the economics of what's possible. A 1M context window has existed before, but filling it was prohibitively expensive. M3 makes long context practical.

The MSA Architecture: Sparse Attention Done Right

The headline technical feature of M3 is MiniMax Sparse Attention (MSA). Standard transformer attention is quadratic—every token attends to every other token, so doubling the context roughly quadruples compute. That's why long context windows have historically been slow and expensive.

MSA replaces full attention with a KV-block selection mechanism. Instead of attending to every token, the model selects the most relevant blocks of the key-value cache for each query. This cuts per-token compute at long context while retaining quality across most tasks. MiniMax reports roughly 9x faster prefill and 15x faster decoding at 1M tokens compared to the prior generation, with per-token compute around one-tenth of previous models at that context length.

This matters because a 1M context window is only useful if you can afford to fill it. MSA turns the long context from a marketing number into a practical tool.

Benchmark Performance: Coding and Beyond

M3 leads with coding and agentic benchmarks. MiniMax reports 59.0% on SWE-Bench Pro (surpassing GPT-5.5 and Gemini 3.1 Pro), 66.0% on Terminal-Bench 2.1 for agentic CLI tasks, and 83.5 on BrowseComp for autonomous browsing—edges past Claude Opus 4.7's 79.3 on that metric.

For an open-weights model you can self-host, closing the gap to the closed frontier on coding is significant. MiniMax also highlighted extreme long-horizon tests: M3 reportedly reproduced the experiments of an ICLR paper in 12 hours and ran continuously for 24 hours while making nearly 2,000 tool calls. These are vendor anecdotes, not standardized benchmarks, but they signal where M3 is aimed: agents that work for hours, not single-turn chat.

The Pricing Gap That Matters

A worked example makes the economics concrete. Say an agentic coding task consumes 500K input tokens and 100K output tokens—a realistic figure once you fill context with repo files and tool output:

• M3 (promo): $0.27 per task
• M3 (standard): $0.54 per task
• Claude Opus: $5.00 per task

At promo pricing, M3 runs the same task at roughly 5% of Opus cost. Even at standard pricing, it's about a tenth. For high-volume agentic workloads, that difference decides whether a product is viable.

How to Access M3

You have three practical paths:

MiniMax Platform API

Create a key at platform.minimax.io and call the OpenAI-compatible endpoint directly. First-party route with full feature support.

OpenRouter

Fastest way to test without a MiniMax account. Point your existing OpenAI-compatible client at OpenRouter and set the model to minimax/minimax-m3. Good for teams already routing across multiple providers.

Self-Host the Open Weights

MiniMax released M3 as open-weights shortly after the API went live. You can run it on your own infrastructure with vLLM or SGLang once they add MSA support. Self-hosting only makes sense at sustained high volume—at low volume, the API is cheaper than idle GPUs. Note the license has commercial-use conditions, so review terms before shipping in a product.

Where M3 Fits in Your Stack

M3 is strongest where its three differentiators line up: cheap long context, agentic tool use, and multimodal input.

Great fits include: long-horizon coding agents over large repos, high-volume agentic workflows where cost dominates, whole-codebase or whole-document analysis, autonomous browsing and research agents, and multimodal pipelines with image and video understanding.

Consider alternatives for: hardest multi-file refactors where Opus still leads marginally, workloads needing strict commercial-use licenses without conditions, latency-critical chat where smaller models suffice, and regulated data requiring specific provider regions.

The pragmatic pattern most teams land on is hybrid: route the bulk of agentic and long-context work to M3 for cost, and reserve a frontier closed model for the small slice of tasks where the last few points of quality matter. A model gateway makes that routing trivial.

Caveats to Keep in Mind

Promo pricing is temporary. Budget against the standard $0.60/$2.40 rate so an expiry doesn't break unit economics. Licensing has conditions—the open weights ship under terms with commercial restrictions that drew criticism at launch. Read the license before building a commercial product on self-hosted M3. Vendor benchmarks are vendor benchmarks—MiniMax's claims about surpassing GPT-5.5 and Gemini 3.1 Pro should be validated on your own workloads. And remember: long context is not memory. A 1M window helps, but stuffing everything into context isn't a substitute for a real memory system on long-running agents.

The Bottom Line

MiniMax M3 represents a shift in what's possible with open-weights models. Frontier coding performance, million-token context, and multimodal input—combined with pricing that makes long-context agents economically viable for the first time. For teams building agentic systems, it's worth testing against your real workloads. The economics alone make it compelling; the performance suggests it might actually deliver.