MiniMax M3: Open-Weight Coding, 1M Context, Multimodal

On May 31, MiniMax quietly released M3 - and it might be the most significant open-weight model drop of the year. Not because it tops any single benchmark, but because it's the first open-weights model to combine three capabilities that have, until now, required frontier closed-source APIs: frontier-level coding, million-token context, and native multimodal understanding.

That's a combination no open model has shipped before. And the pricing - $0.30 per million input tokens on OpenRouter - makes it one of the cheapest frontier-adjacent models available.

What M3 Actually Delivers

MiniMax M3 is a multimodal foundation model that accepts text, image, and video inputs with text output. It uses MiniMax's proprietary Sparse Attention (MSA) architecture, which they claim delivers 9.7× faster prefilling and 15.6× faster decoding at 1M tokens compared to standard attention. The context window is 1M tokens on the API, with a guaranteed minimum of 512K.

The headline specs:

• Multimodal from the ground up - Not a text model with a vision adapter bolted on. MiniMax says they restructured the entire data pipeline to hundreds of terabytes and trained multimodal from step zero, aligning text and visual semantic spaces natively.
• 1M token context - The same class as Gemini Ultra 2 and Grok 4.3, achieved through the MSA architecture rather than brute-force sequence length.
• Frontier coding and agentic capability - MiniMax claims M3 is the first open model to combine all three: coding, long context, and multimodal - a trifecta that's previously been the exclusive domain of models like Claude Opus 4.8 and GPT-5.5.

The Benchmark Claims

MiniMax's benchmark numbers are impressive, though they come with the usual caveat of self-reported results:

• BrowseComp: 83.5 - Surpassing Claude Opus 4.7's 79.3 on this agentic browsing and information retrieval benchmark. This is a notable result - BrowseComp measures the kind of autonomous web navigation that's increasingly important for real-world agent workflows.
• PostTrainBench: 37.1 - M3 ranked third behind Opus 4.7 (42.4) and GPT-5.5 (39.3) on this benchmark that tests a model's ability to autonomously train and improve other models. That's a remarkable result for an open-weights model.
• 12-hour autonomous paper reproduction - MiniMax demoed M3 independently reproducing an ICLR 2025 Outstanding Paper over 12 hours, producing 18 commits and 23 experiment charts with zero human intervention.
• CUDA kernel optimization - M3 spent 24 hours optimizing an FP8 matrix multiply kernel on NVIDIA Hopper, running 147 benchmark iterations and 1,959 tool calls to achieve a 9.4× speedup (from 7.6% to 71.3% hardware utilization), again with zero human intervention.

These aren't cherry-picked quick tests. They're long-horizon agentic runs that stress-test exactly the kind of sustained, autonomous work that separates frontier models from the rest.

How It Compares on the LMRank Leaderboard

M3 hasn't completed independent evaluation on our methodology yet, but we can contextualize it against the models it's competing with:

Model	LMRank	Input $/M	Output $/M	Context	Multimodal
Claude Opus 4.8	9.7	$15.00	$75.00	200K	Yes
GPT-5.5	9.4	$10.00	$30.00	256K	Yes
Gemini Ultra 2	9.2	$5.00	$15.00	2M	Yes
Qwen3.7 Max	9.0	$2.50	$10.00	128K	No
MiniMax M3	TBD	$0.30	$1.20	1M	Yes
DeepSeek V4	9.0	$0.27	$1.10	128K	No

The pricing story is stark. M3 costs roughly 1/50th of Claude Opus 4.8 per million input tokens while claiming to compete in the same capability class. Even compared to DeepSeek V4 (9.0, $0.27/M input), M3 offers a much larger context window and native multimodal at nearly the same price.

The Open-Weights Factor

MiniMax says M3 will be fully open-sourced on HuggingFace and GitHub, supporting private cluster deployment and fine-tuning. That's a meaningful differentiator. While Qwen3.7 Max is API-only and Gemini Ultra 2 is locked to Google's infrastructure, M3 aims to deliver comparable capabilities with full deployment flexibility.

The implications for enterprise adoption are significant. Organizations that can't send data to external APIs - healthcare, finance, government - now have a viable path to frontier-class multimodal AI on their own hardware. The 1M token context means entire codebases, document libraries, or video archives can be processed in a single pass.

What to Watch For

Self-reported benchmarks are starting points, not conclusions. Here's what matters next:

• Independent evaluation - LiveBench, LMSYS Arena, and third-party SWE-Bench runs will tell us whether M3's claims hold up under blind testing.
• Open-weight release quality - The gap between "available on API" and "actually usable as open weights" can be enormous. We'll be watching for model card completeness, quantization support, and community adoption speed.
• Real-world agentic performance - The 12-hour paper reproduction and CUDA optimization demos are impressive, but they're controlled experiments. How does M3 perform in messy, real-world coding agents with ambiguous requirements and shifting context?
• The Chinese open-source pipeline - M3 joins DeepSeek V4, Qwen3.7 Max, and Qwen3.6 35B A3B in a rapidly growing roster of competitive Chinese open-weight models. The pace of releases suggests this pipeline isn't slowing down.

The Bottom Line

MiniMax M3 isn't the model that dethrones Claude Opus 4.8 at the top of the leaderboard. But it might be the model that makes the top of the leaderboard irrelevant for most use cases. When you can get frontier-level coding, million-token context, and native multimodal for $0.30 per million input tokens - with open weights for self-hosting - the value equation shifts fundamentally.

The real test isn't whether M3 beats Opus on a benchmark. It's whether M3 is good enough, at its price point, to make organizations stop paying 50× more for marginal gains. Early signs suggest it might be.

At lmrank.com, we track live scores and pricing for every major LLM. Follow the leaderboard for updated rankings as MiniMax M3 and other new models complete independent evaluation.

See also: Best Open-Weight Models