MiniMax M3 turns long-context AI into an agent platform test
MiniMax launched M3 on June 1, 2026, combining long-context, agentic, coding and native multimodal capabilities in one model line. The API supports up to 1 million tokens of context, with a guaranteed minimum of 512K tokens, and includes M3 and M3-highspeed versions. MiniMax plans to open-source M3 on HuggingFace and GitHub, while early pricing offers a 50% discount for the first seven days.
MiniMax has introduced M3, a new flagship AI model that puts long-context reasoning, agentic workflows, coding performance and multimodal processing into one product line.
The Shanghai-based company, also known as Shanghai Hixi Technology, launched the model on June 1, 2026.
What happened
M3 is being positioned as a domestic AI model built for demanding agent and coding tasks.
Its API supports a context window of up to 1 million tokens, with at least 512K tokens guaranteed, and the model is trained as natively multimodal rather than as a text-only system with later visual additions.
MiniMax says the model uses its proprietary Sparse Attention architecture.
The company also says it rebuilt its data pipeline, expanded pre-training data to hundreds of terabytes, and worked on alignment between text and visual semantic spaces.
The launch includes two API versions: M3 and M3-highspeed.
Both are described as producing identical results, while M3-highspeed is designed for faster inference.
Automatic caching is enabled by default.
Why it matters
The announcement is a signal that Chinese AI model developers are moving beyond headline benchmark competition and into agent infrastructure.
A model with a million-token window could matter for developers building long-running coding sessions, research assistants, document-heavy workflows or video-understanding tools.
The key commercial question is whether these capabilities translate into reliable enterprise use.
Long context, tool use and autonomous task execution may reduce friction for teams experimenting with AI agents, but adoption would depend on stability, cost and the quality of downstream applications.
Performance signals
MiniMax highlighted several benchmark and demonstration results.
In BrowseComp, M3 scored 83.5, compared with 79.3 for OpenAI Opus 4.7.
In one autonomous experiment, M3 spent nearly 12 hours reproducing an ICLR 2025 outstanding paper on LLM fine-tuning dynamics, producing 18 commits and 23 experimental charts.
The company also tested M3 as an AI research assistant.
In that task, the model was given four pre-trained base models and asked to carry out data synthesis, training, evaluation and iteration within 12 hours without human intervention.
M3 scored 37.1, behind Opus 4.7 at 42.4 and GPT-5.5 at 39.3.
What to watch next
MiniMax plans to open-source M3 on HuggingFace and GitHub, with support for private cluster deployment and fine-tuning.
That could make the model more relevant to teams that want more control over infrastructure and customization.
Pricing will also shape market response.
For the first seven days, MiniMax is offering a 50% discount for M3 API usage at contexts up to 512K tokens, with input, output and cache-read pricing set across standard and priority tiers.
Readers should watch whether developers treat M3 as a practical agent platform rather than only a benchmark announcement.
