MiniMax M3: The New Frontier of Open-Weights Models and the MSA Architecture

MiniMax M3: The New Frontier of Open-Weights Models and the MSA Architecture

For the past two years, we as software developers have watched the same scene play out. Closed-source models sweep the benchmark tables, and the open-source community scrambles to close the gap. On June 1, 2026, MiniMax M3 broke that dynamic. Xiyu Technology's (MiniMax) new flagship release for the open-source world ships with a 1 million token context window, native multimodality, and a brand-new attention mechanism called MiniMax Sparse Attention (MSA). On top of that, it surpasses GPT-5.5 and Gemini 3.1 Pro on coding and agentic tasks while approaching Claude Opus 4.7.

This article walks through M3's architecture, benchmark results, pricing, access paths, and what it really means for the developer world. No fluff, just concrete numbers and examples.

1. Why Is M3 Making So Much Noise?

Frontier models are now defined by three capabilities: top-tier coding, very long context windows, and native multimodality. An open-weights model that combines all three did not exist until now. M3 is the first open-weights package to bring them together. As highlighted in the official MiniMax blog post, these three features have become "table stakes" for any serious frontier contender.

Here is M3's quick identity card:

• Release date: June 1, 2026 (API went live May 31)
• Architecture: Sparse Mixture-of-Experts with MiniMax Sparse Attention (MSA)
• Context window: 1 million tokens (guaranteed 512K minimum)
• Modalities: Text, image, and video in; text out
• Weights: Open-weights (released shortly after the API launch)
• Best at: Coding, agentic tool use, long-context reasoning, multimodal

M3 looks like a successor to MiniMax's M2 family (M2, M2.5, M2.7), but it is a generational leap. The M2 series had dropped sparse attention in favor of full attention. M3 brings sparse attention back in a new form. This is the newest link in the open-source coding model wave we covered in our Kimi K2.6 article.

2. MSA: A New Form of Sparse Attention

Standard transformer attention carries quadratic cost. Every token attends to every other token in the context. Doubling the context quadruples the attention compute. That is why long context windows have been slow and expensive for years. We explored a similar bottleneck in our FlashQLA article, where the Qwen team built a linear attention kernel to push agentic AI to edge devices.

MSA replaces full attention with a KV-block selection mechanism. For each query, the model selects the most relevant key-value cache blocks. The result is lower per-token compute at long context while preserving quality. The gains reported in the official announcement are striking:

• ~9x faster prefill at 1M tokens, compared to the previous generation
• ~15x faster decoding at 1M tokens, compared to the previous generation
• Per-token compute at 1M tokens is just 1/20 of the previous model
• 4x faster than open-source Flash-Sparse-Attention and flash-moba at the operator level

MSA differs from competitors like DSA and MoBA by partitioning the KV cache into blocks more precisely. At the operator level, it adopts a "KV outer gather Q" approach: each block is read only once, memory access is contiguous, and arithmetic intensity is dramatically higher than common methods. All these optimizations turn theoretical gains into practical reality.

3. Benchmarks: Coding, Agentic, and Multimodal

M3's boldest claims are in coding. Let us look at SWE-Bench Pro, Terminal-Bench 2.1, SWE-fficiency, KernelBench Hard, and MCP Atlas:

• SWE-Bench Pro: 59.0% — long-horizon real-world software engineering tasks (surpasses GPT-5.5 and Gemini 3.1 Pro, approaches Opus 4.7)
• Terminal-Bench 2.1: 66.0% — agentic terminal and CLI task completion
• SWE-fficiency: 34.8% — efficient resolution of engineering tasks
• KernelBench Hard: 28.8% — GPU kernel optimization
• MCP Atlas: 74.2% — Model Context Protocol tool use

On BrowseComp, also covered in the Lushbinary developer guide, M3 scores 83.5, beating Opus 4.7's 79.3. On OmniDocBench for multimodal, it climbs above Gemini 3.1 Pro. On Claw-Eval for end-to-end agent evaluation, it takes the top score.

A caveat: vendor-published benchmarks always deserve healthy skepticism. Before you commit M3 to production, run it against a held-out evaluation set built from your own real tickets, prompts, and tool traces. The eval-driven approach we discussed in our Karpathy Autoresearch article applies here as well.

4. Real-World Tasks: 12 and 24 Hours of Autonomous Work

Benchmarks alone are not enough. M3's most interesting side is its performance on long-horizon agent tasks. The MiniMax team shared three striking real-world scenarios:

Scenario 1: Independent Paper Reproduction. The team gave M3 an ICLR 2025 Outstanding Paper Award winner, "Learning Dynamics of LLM Finetuning." M3 ran autonomously for about 12 hours, produced 18 commits and 23 experimental figures, and successfully matched the SFT prediction-probability change trend. It observed the "squeezing effect" in the DPO experiments and verified the paper's proposed Extend mitigation method. Doing this required fitting the paper, code, and experiment logs into context at once.

Scenario 2: CUDA Kernel Optimization. Writing a production-grade FP8 GEMM kernel on NVIDIA Hopper architecture typically takes an experienced team 1-2 weeks. M3 received only a task description, a benchmark evaluation script, and a non-runnable Triton skeleton. No reference high-performance implementation was available. Over roughly 24 hours of continuous execution, M3 completed 147 benchmark submissions and 1,959 tool calls. It took Hopper FP8 hardware peak utilization from 7.6% in version 1 to 71.3%, a 9.4x speedup. The best solution appeared on the 145th submission, after passing through multiple performance plateaus. This shows M3 is an agent that does not give up. We saw a similar "smart tool, durable agent" mindset in our Zed 1.0 editor article.

Scenario 3: Training Base Models (PostTrainBench). The team gave M3 four Base models that had only completed pretraining. Within 12 hours, M3 autonomously completed the entire process of data synthesis, training, evaluation, and iteration without human intervention. Final score: 0.37 across AIME2025, BFCL, GPQA Main, GSM8K, and HumanEval, slightly below Opus 4.7 (0.42) and GPT-5.5 (0.39), but clearly ahead of the rest.

5. Pricing: The Open-Source Economic Edge

M3 was listed on OpenRouter with a temporary 50% promotion at $0.30 per million input tokens and $1.20 per million output tokens. The standard rate is $0.60 / $2.40. For comparison:

• M3 (promo): $0.30 input / $1.20 output
• M3 (standard): $0.60 input / $2.40 output
• Claude Opus 4.x: $5.00 input / $25.00 output
• GPT-5.5: ~$10.00 input / ~$30.00 output

A worked example: an agentic coding task consuming 500K input + 100K output tokens (realistic once you fill a long context with repo files and tool output):

• M3 (promo): (0.5 × $0.30) + (0.1 × $1.20) = $0.27 per task
• M3 (standard): (0.5 × $0.60) + (0.1 × $2.40) = $0.54 per task
• Claude Opus: (0.5 × $5.00) + (0.1 × $25.00) = $5.00 per task

At promo pricing, M3 runs the same task at roughly 5% of Opus's cost, and even at standard pricing it is about one-tenth. For high-volume agentic workloads, this difference decides whether a product is viable. M3 takes the efficient-AI trend we covered in our SLM article to the high end.

On the subscription side, MiniMax Token Plan comes in three tiers: Plus ($20/month, ~1.7B M3 tokens), Max ($50/month, ~5.1B M3 tokens), and Ultra ($120/month, ~9.8B M3 tokens). The same token pool covers text, image, speech, and music generation. Among comparable plans, this is one of the highest token quotas globally.

6. MiniMax Code: The Agent Product Trained Alongside M3

Alongside M3, MiniMax Code was updated. This agent product, trained together with M3, fully exploits M3's long context, coding/agentic capabilities, and native multimodality. The Agent Team feature breaks large tasks into multi-stage, concurrent, and dynamically adjustable workflows. Through a Producer + Verifier adversarial loop, it produces, reflects, and corrects itself continuously. It can run autonomously for days without human intervention.

MiniMax Code moves in a similar direction to Claude Code's recent Dynamic Workflows release, but instead of fixed JS-based orchestration, it focuses on "deep reflection and continuous error correction": the agent adjusts its plans and priorities in real time based on task progress, while users can step in at any time to add requirements or correct the direction. Thanks to native multimodality, computer use is supported too. From your phone, you can say "Open the local ERP client and batch-enter invoice information based on this Excel spreadsheet," and M3 completes the operation across applications on your behalf.

Notably, MiniMax Code is built on a harness based on the open-source community projects OpenCode and Pi. MiniMax plans to open-source the project in the future. This is an important signal for the developer community.

7. API and Access Paths

You can reach M3 through three practical paths:

A. MiniMax Platform API: Get an API key from platform.minimax.io and use the OpenAI-compatible endpoint:

curl https://api.minimax.io/v1/chat/completions \\
  -H "Authorization: Bearer *** \\
  -H "Content-Type: application/json" \\
  -d '{
    "model": "MiniMax-M3",
    "messages": [
      {"role": "user", "content": "Summarize this repo and propose a refactor plan."}
    ]
  }'

B. OpenRouter: The fastest path to test M3 at promo pricing. Point your OpenAI-compatible client at OpenRouter and set the model name to "minimax/minimax-m3."

C. Self-Host the Open Weights: Once the open weights are released, you can run M3 on your own infrastructure with inference engines like vLLM or SGLang once they add MSA support. Self-hosting only makes sense at sustained high volume; at low volume, the API is far cheaper than idle GPUs. Note that the license carries commercial-use conditions, so review the terms before shipping it in a product.

On the API side, two service tiers are available: the default standard tier for regular requests, and the priority tier (service_tier=priority) for SLA-sensitive industrial use cases that need scheduling priority and stable response latency under high concurrency. The priority channel is currently enabled through sales support and is expected to open to all users within days.

8. Where to Use M3 (and Where Not)

When M3's three differentiators line up, its sweet spots become clear:

Great fit:

• Long-horizon coding agents over large repositories
• High-volume agentic workflows where cost dominates
• Whole-codebase or whole-document analysis
• Autonomous browsing and research agents
• Multimodal pipelines (image, video understanding)

Consider alternatives:

• Hardest multi-file refactors where Opus still leads marginally
• Workloads needing a license without commercial-use conditions
• Latency-critical chat where smaller models suffice
• Regulated data requiring a specific provider region

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 gateway makes that routing trivial.

9. Caveats to Watch

Do not overlook a few important points:

• Promo pricing is temporary. Budget against the standard $0.60/$2.40 rate so a promo expiry does not break your unit economics.
• The license has conditions. The open weights ship under terms with commercial restrictions. Read the license before building a commercial product on self-hosted M3.
• Vendor benchmarks are vendor benchmarks. The "surpasses GPT-5.5 and Gemini 3.1 Pro" claims are MiniMax's own. Independent leaderboards and your own evals are the real test.
• Long context is not memory. A 1M window helps, but stuffing everything into context is not a substitute for a real memory system on long-running agents. The observability principles we covered in our AI-era logging article apply here: without logging agent traces, tool calls, and decisions, debugging long-horizon tasks becomes impossible.

10. Conclusion: A New Frontier for Open Source

MiniMax M3 is a significant turning point in the 2026 open-source LLM landscape. A 1M token context window, MSA-driven efficiency, native multimodality, and near-frontier performance in coding and agents make it a serious production candidate. As an open-weights model with a major price advantage, it is especially attractive for high-volume agentic workloads.

However, rather than betting your entire workload on a single model, a hybrid strategy is wiser. M3 for cost-critical jobs, Opus or GPT-5.5 for the few tasks where the last few quality points matter. Setting up that routing through a gateway protects both development speed and unit economics.

The fastest way to try M3 on your own workload is to grab an API key from the MiniMax platform and test it through OpenRouter at the promo rate. Compare the results against your own evaluation set. In a period when open source has matured this much, staying flexible across models, rather than locking into a single one, is the biggest competitive advantage.

---

This article was prepared with the assistance of the MiniMax M3 model itself. The model's MSA architecture, benchmark results, pricing, and real-world usage scenarios were compiled from the official MiniMax announcement and the Lushbinary developer guide.