MiniMax released their latest flagship model M3 on June 1, 2026, and it represents a significant milestone in the open-weight LLM space. M3 is the first and only open-weight model to combine three capabilities that have been table stakes for closed-source frontier models: frontier-level coding and agentic work, a 1 million token context window, and native multimodal input (image and video).
While the closed-source frontier — OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7 — continues to push the envelope, M3 is the only open model that brings all of these together in a single package. And its benchmark scores are genuinely competitive.
What Makes M3 Different: MSA Architecture
The most significant architectural innovation in M3 is MiniMax Sparse Attention (MSA) — a new sparse attention mechanism designed specifically to solve the "inherent flaw" of full attention: quadratic computational complexity growth.
Most sparse attention mechanisms avoid complexity explosion by adding a pre-filtering stage. MSA goes further by partitioning the KV cache into blocks more precisely than approaches like DSA and MoBA, achieving higher effective context coverage. At the operator level, MSA adopts a "KV outer gather Q" approach that uses KV blocks as the outer loop to aggregate queries that hit them.
MSA Performance Numbers
Each KV block is read only once with contiguous memory access. MSA is more than 4x faster than open-source Flash-Sparse-Attention and flash-moba. At a context length of 1 million, M3's per-token compute is just 1/20 that of the previous-generation model, achieving 9x speedup in the prefilling stage and 15x speedup in decoding.
This isn't just an incremental improvement — it's a fundamental shift in how long-context inference works at scale. The clean, scalable, and hardware-friendly design means theoretical gains are fully realized in practice.
Frontier Coding Benchmarks
M3 shows significant improvements over its predecessor M2, approaching the level of leading closed-source models in bugfix, frontend/backend development, and performance optimization. Here are the key benchmark scores:
| Benchmark | M3 Score | Significance |
|---|---|---|
| SWE-Bench Pro | 59.0% | Software engineering tasks |
| Terminal-Bench 2.1 | 66.0% | Terminal/CLI execution |
| SWE-fficiency | 34.8% | Coding efficiency |
| KernelBench Hard | 28.8% | Kernel-level optimization |
| MCP Atlas | 74.2% | Tool/scaffold generalization |
What's notable is M3's Terminal-Bench 2.1 score of 66% — this puts it in contention with closed-source leaders on terminal-based coding tasks. For an open-weight model, this is genuinely impressive.
Interactive User Simulator for Coding Training
One of M3's most innovative training approaches is MiniMax's interactive user simulator framework. Most coding benchmarks evaluate models on single-turn tasks, but real developers don't work that way. They clarify requirements, adjust solutions, assign tasks across contexts, and iterate over multiple rounds.
The user simulator exposes M3 to interaction scenarios closer to production environments — simulating requirement elaboration, solution discussion, feedback-based correction, continuous task switching, and complex project iteration. This means M3 doesn't just execute instructions passively; it can actively collaborate with users.
Why This Matters
The next generation of agentic coding won't be measured only by code generation scores. It'll be measured by long-term collaboration capability, planning ability, and efficiency of human-agent workflows. M3 was trained with this philosophy from the ground up.
Agentic Capabilities
M3 isn't just a coding model — it's a general-purpose agentic model that performs strongly on office workflows, search tasks, and even financial domain tasks.
The most impressive demonstrations come from real-world tasks MiniMax tested internally:
- 12-Hour Autonomous Paper Reproduction: M3 independently reproduced an ICLR 2025 Outstanding Paper ("Learning Dynamics of LLM Finetuning") — running for nearly 12 hours, producing 18 commits and 23 experimental figures. It parsed charts and formulas from the paper, fit the entire paper + code + experiment logs into a single 1M context window, and used its coding + agentic capabilities to drive long-horizon execution.
- CUDA Kernel Optimization: M3 optimized an FP8 GEMM kernel on NVIDIA Hopper GPUs — starting from only a task description and a non-runnable Triton skeleton. Over ~24 hours, it completed 147 benchmark submissions and 1,959 tool calls, pushing hardware peak utilization from 7.6% to 71.3% — a 9.4x speedup with zero human intervention.
- PostTrainBench: Training Models on Its Own: M3 autonomously completed a full pipeline — data synthesis, training, evaluation, and iteration — across four pretrain-only base models within 12 hours. It scored 37.1, ranking #3 overall, behind only Claude Opus 4.7 (42.4) and GPT-5.5 (39.3), significantly ahead of all other models.
Native Multimodal: Image, Video & Desktop Control
M3 is a natively multimodal model that was trained with mixed-modality data from Step 0. This means the semantic spaces of different modalities merge more naturally than in models that bolted multimodal capabilities onto a text-only foundation later.
Key multimodal capabilities include:
- Image input — understanding and reasoning over visual content
- Video input — processing temporal visual information
- Desktop computer operation — the ability to navigate and control a desktop environment
- Interleaved data training — MiniMax re-architected their entire text pretraining pipeline to produce large volumes of interleaved (text + image/video) training data, finding that interleaved data scales more easily than synthetic data
The desktop control capability is particularly interesting from an agentic standpoint — it means M3 can interact with any desktop application as a tool, expanding its reach far beyond terminal and API environments.
How to Access M3
MiniMax offers multiple ways to access M3:
- API: Full API access with automatic cache support, no configuration needed. Model identifier:
MiniMax-M3 - MiniMax Code: Direct integration with MiniMax's AI coding environment
- Token Plan: Subscription-based access with invite discounts (10% for both inviter and invitee)
- AI Coding Tools: Compatible with Claude Code, Roo Code, Kilo Code, Cline, Codex CLI, OpenCode, Droid, TRAE, Grok CLI, and Cursor
Open Weight
M3 is the first open-weight model combining all three capabilities (coding, agentic, long-context, multimodal). Being open-weight means researchers and developers can inspect, fine-tune, and deploy it locally — a significant advantage for privacy-conscious deployments and custom model training.
MiniMax: The Company Behind M3
MiniMax is a Shanghai-based AI company founded in early 2022. Their mission is "Intelligence with Everyone" — building AGI accessible to everyone. Key company stats:
- 200M+ global individual users across 200+ countries and regions
- 1M+ enterprise clients and developers across 100+ countries
- Founded by Yan Junjie (CEO), Yang Bin, and Zhou Yucong
- Traded on SEHK as stock 100
- Product portfolio includes MiniMax M3, M2.7, M2.5 (LLMs), Hailuo 2.3 (video), Speech 2.8 (speech), Music 2.6 (music), MiniMax Code, and their enterprise API platform
MiniMax is one of China's most prominent AI companies, competing directly with DeepSeek, Alibaba's Qwen, and ByteDance's models. M3 represents their push into the open-weight frontier space — a space where they compete with Meta's Llama family and the growing number of open-weight frontier models.
Verdict
MiniMax M3 is a genuinely impressive open-weight model. Its combination of frontier coding performance, agentic capability, native multimodality, and 1M context window in a single model is unique in the open-weight space. The MSA sparse attention architecture is a real architectural contribution that could influence the broader field.
For developers who need a powerful coding agent with long-context awareness and multimodal understanding — without being locked into a closed ecosystem — M3 is worth serious consideration. The fact that it can independently reproduce research papers, optimize CUDA kernels for 24 hours straight, and even train other models from scratch suggests we're entering a new era of autonomous AI capability.