MiniMax-M3

MiniMax-M3 is an open native multimodal model with support for context up to 1 million tokens, combining advanced programming, agentic workflow, and native multimodal capabilities. The model features a Mixture of Experts (MoE) architecture with ~428 billion total parameters, of which only ~23 billion are activated at each generation step, 128 local experts with top-4 expert activation per token, and one shared expert. This sparsity delivers high performance with efficient compute resource usage. M3 supports three reasoning (thinking) modes: *enabled* — reasoning always active, *adaptive* — the model automatically determines whether additional reasoning is required, and *disabled* — reasoning turned off to minimize latency. In contrast to the previous MiniMax-M2.7 model and the entire second-generation family, MiniMax-M3 represents a qualitative step forward by combining three cutting-edge capabilities — native multimodality, a million-token context length, and outstanding performance on programming and agentic tasks.

The key architectural innovation in MiniMax-M3 is MiniMax Sparse Attention (MSA) — a new type of sparse attention specifically designed for efficient processing of contexts up to one million tokens. MSA is built on top of Grouped Query Attention (GQA) and consists of two parallel processes: a lightweight Index Branch evaluates all key-value blocks in the input sequence and independently selects the most relevant subset of blocks for each query group; the main Main Branch then performs exact block-sparse attention exclusively over the selected blocks. The local block (the immediate neighborhood) is always included regardless of its score, ensuring that important contextual information is preserved. Thanks to this approach, MSA reduces per-token attention computation costs by 28.4× at a context length of 1 million tokens compared to GQA.

Results on key benchmarks confirm the quality of the M3 model. On SWE-bench Verified (the reference benchmark for evaluating AI’s ability to solve real-world tasks from software repositories) the model achieves 80.5% resolved tasks. On SWE-bench Pro the score is 59%. On BrowseComp (a benchmark for evaluating autonomous web browsing and information extraction) M3 scores 83.5 points, outperforming Opus 4.7 (79.3). In the autonomous PostTrainBench test, where the model was tasked with autonomously conducting a full post-training cycle of other models (data synthesis, training, evaluation, iterations) within 12 hours, M3 placed 3rd with a result of 37.1, behind only the closed models Opus 4.7 (42.4) and GPT-5.5 (39.3), and significantly ahead of all other models.

The use cases for MiniMax-M3 are extremely broad and span many domains where advanced agentic and multimodal capabilities are required. In software development, the model can act as an autonomous AI assistant capable of analyzing entire code repositories, performing long-running refactoring, debugging, and documentation tasks, as well as autonomously deploying and testing code. Thanks to the 1-million-token context window, M3 is ideally suited for analyzing and understanding long documents, including scientific papers with charts and formulas, legal contracts, financial reports, and complete system log files. In multimodal applications, the model can be used for video understanding, extracting information from complex visual materials, and building interactive systems that work with various data types. Agentic scenarios include autonomous execution of research tasks, tool and API management, and creating systems capable of independently planning and executing multi-step workflows without human intervention.

Announce Date: 02.06.2026
Parameters: 428B
Experts: 128
Activated at inference: 23B
Context: 1049K
Layers: 60
Attention Type: MiniMax Sparse Attention (MSA)
Developer: MiniMax-AI
Transformers Version: 4.52.4
vLLM Version: >=0.24.0
License: MINIMAX COMMUNITY LICENSE

Use our pre-built public endpoints for free to test inference and explore MiniMax-M3 capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.