MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model
Introduction
On June 1, 2026, MiniMax officially released its new flagship large language model, MiniMax M3. Positioned as "Coding & Agentic Frontier · 1M-context MSA · Native Multimodality," M3 is a full-stack model that combines three of the most demanding capabilities in modern LLMs. At a time when overseas closed-source frontier models have long treated coding, agentic work, and long context as table stakes, M3 is the first model in China to bring all three together — and the only open-weight model in the world to do so today. This article takes a deep dive into why M3 deserves the tagline "All in One Model," covering its architecture, capabilities, benchmarks, and product surface.
1. Three Frontier Capabilities, One Open-Weight Model
For a long time, a quiet capability gap has separated open-weight LLMs from closed-source frontier models: the latter have continued to lead on coding, agentic behavior, and long context, while open-weight models have typically chased one or two of these dimensions. M3 breaks that pattern.
- Frontier Coding: Compared with its predecessor M2, M3 shows significant improvement on bug fixes, frontend/backend development, and performance optimization, approaching the level of leading overseas closed-source models.
- 1M Ultra-Long Context: The API supports a context window of up to 1M tokens, with a guaranteed usable length of at least 512K tokens. It is the infrastructure foundation for long-horizon agents, long-horizon coding, and long-video understanding.
- Native Multimodality: M3 supports image and video input and can operate a desktop computer (Computer Use) — a genuine "see images, see video, control the desktop" stack.
These three capabilities are the entry ticket for closed-source frontier models today. M3 is the first to bring that combination to the open-weight world.
2. MSA: A New Sparse Attention that Makes Context a Scalable Dimension
To truly deliver 1M context, the model has to start at the most fundamental layer — the attention mechanism — and avoid the "inherent flaw" of full attention: quadratic growth of computational complexity. M3's answer is MSA (MiniMax Sparse Attention) — a clean, extensible, and hardware-friendly sparse attention architecture.
2.1 Architectural Design
Sparse attention mechanisms generally avoid the complexity-explosion problem by adding a pre-filtering stage. Compared with approaches such as DSA and MoBA, MSA can partition the KV cache into blocks more precisely, achieving higher effective context coverage. MSA is not a quick engineering patch but a structural redesign of the attention pipeline.
2.2 Operator-Level Optimization
MSA optimizes directly at the operator level with a "KV outer gather Q" approach: KV blocks act as the outer loop, and the queries that hit each block are aggregated. Each block is read only once, and memory access is contiguous. Under M3's head configuration, arithmetic intensity is significantly better than common methods — more than 4× faster than the open-source Flash-Sparse-Attention and flash-moba implementations.
2.3 Real-World Gains
The clean, scalable, easy-to-implement, and hardware-friendly design of MSA means its theoretical benefits translate into production performance:
- At a 1M-token context length, M3's per-token compute is just 1/20 of the previous-generation model.
- The prefilling stage sees a speedup of more than 9×.
- The decoding stage sees a speedup of more than 15×.
- Across multiple ablations, MSA matches full attention on the vast majority of capabilities, with no significant quality loss.
In other words, MSA uses 1/20 of the compute to deliver 1M context at comparable quality — this is what true "Context Scaling" looks like.
3. Benchmarks: M3 Reaches the International Frontier on Coding & Agentic Tasks
Coding and agentic capabilities are the headline improvements of M3. Across internationally recognized benchmarks covering software engineering, terminal execution, and protocol-based tool use, M3 reaches the frontier:
| Benchmark | Domain | M3 Score |
|---|---|---|
| SWE-Bench Pro | Real-world GitHub issue resolution | 59.0% |
| Terminal-Bench 2.1 | Terminal command execution | 66.0% |
| SWE-fficiency | Software engineering efficiency | 34.8% |
| KernelBench Hard | Kernel / system-level code | 28.8% |
| MCP Atlas | MCP protocol tool use | 74.2% |
M3's 74.2% on MCP Atlas is especially notable — it means the model can stably operate in complex tool-use scenarios orchestrated by the Model Context Protocol. Its 59.0% on SWE-Bench Pro also makes M3 the first open-weight model in China to enter the top tier on this benchmark.
4. Interactive User Simulator: From Passive Execution to Active Collaboration
The M3 team surfaces a long-overlooked issue: most current training and evaluation of code agents assume single-turn tasks. But real usage is not like that. Users collaborate continuously within the same session — clarifying requirements, adjusting solutions, dispatching tasks across contexts, and iterating over multiple rounds based on intermediate results.
To close the gap between benchmarks and real user experience, MiniMax built an interactive user simulator framework. By simulating the behavioral patterns of real developers during collaboration, the framework exposes the model during both training and evaluation to interaction scenarios that are much closer to production environments. It can simulate requirement elaboration, solution discussion, feedback-based correction, continuous task switching, and complex project iteration.
The next generation of agentic coding is no longer measured purely by code generation, but by long-term collaboration capability, planning ability, and the efficiency of human-agent collaboration. M3 scales up the data that truly matters for coding and agents. Its goal is not only to lead on benchmarks, but to become a genuinely useful daily partner for developers.
5. Product Surface and Ecosystem
M3 is available to developers and enterprise users through three entry points:
- MiniMax Code (code.minimaxi.com): the official AI coding product, offering first-hand access to M3's coding and agentic capabilities.
- Token Plan: a subscription-based package for individual developers, providing metered access to the M3 API.
- API services: available via platform.minimaxi.com, billed per token, and ready to power third-party applications.
M3's open-weight release means that, for the first time, Chinese developers and enterprises can deploy a flagship model with frontier coding, 1M context, and native multimodality on their own infrastructure — without depending entirely on closed-source APIs.
Conclusion
The release of MiniMax M3 marks the beginning of a new phase: open-weight models have, for the first time, fully stood on the same stage as closed-source frontier models across the three most critical dimensions — coding, long context, and multimodality. The MSA architecture opens the door to 1M context with just 1/20 of the compute; the interactive user simulator teaches agents how to truly collaborate; and the benchmark results give M3 the most accurate label in the industry: "the first open-weight flagship in China to unite all three frontier capabilities." Over the next 12 months, the combination of long context and coding agents will become the central battlefield for LLM application deployment.
Sources: MiniMax M3 official release report (CN) · MiniMax M3 model page · M3 release report (EN)