MiniMax Unveils M3 Large Model: Focusing on Full-Stack Agents with Breakthroughs in Performance and Efficiency
Chinese AI company MiniMax recently released its M3 series of large models, with plans to open-source them in the future. According to its technical report and initial tests, M3 demonstrates outstanding performance in areas like software engineering, multimodal understanding, and autonomous agent evaluation. It aims to solve the “impossible triangle” of AI Agent development: the challenge of simultaneously achieving strong coding capabilities, an ultra-long context window, and stable tool-calling abilities.
Outstanding Performance in Key Benchmarks
M3 has achieved leading scores in a series of industry-standard benchmarks, showcasing its competitiveness across multiple dimensions.
- Software Engineering Capabilities: On the SWE-Bench Pro test set, which measures a model’s ability to solve real-world software engineering problems, M3 scored 59.0, surpassing GPT-5.5 and Gemini 3.1 Pro, and approaching the performance of Opus 4.7.
- Terminal Programming: In terminal operation tasks, M3’s score is on par with Opus 4.7.
- Multimodal Understanding: On the OmniDocBench benchmark for multimodal document understanding, M3’s score exceeded that of Gemini 3.1 Pro.
- Autonomous Agent Evaluation: On Claw-Eval, an end-to-end autonomous agent evaluation framework, M3 achieved the highest score among currently published models.
These results indicate that M3 has been comprehensively optimized for the core capabilities required for AI Agents to execute complex tasks.
MSA Architecture and “Dynamic Workflows” Boost Efficiency
M3 introduces a new MSA (Mixture-of-Searching-and-Attention) architecture, significantly improving the model’s efficiency in processing long contexts. According to official documentation, this architecture reduces the computational load for processing million-token contexts to 1/20th of the previous generation’s models. Specifically, the prefill speed is increased by 9x, and the decoding speed by 15x. This technological breakthrough allows AI Agents to understand context and generate responses more quickly when analyzing large codebases or processing massive documents.
At the application level, M3 supports “Dynamic Workflows,” allowing users to initiate parallel multi-agent task processing via specific commands (e.g., /workflow) or mode switching (e.g., auto mode). This workflow can break down complex tasks to be collaboratively completed by multiple sub-agents. Users can preview the execution script in real-time and intervene during the process.
Case Studies: Full-Stack Agent in Practice
To validate M3’s capabilities in real-world scenarios, developers conducted tests on two complex software engineering tasks.
Case Study 1: Automating the Training of a Small GPT Model
This test aimed to replicate a project similar to Andrej Karpathy’s nanoGPT, covering the entire pipeline from data preparation and tokenization to model pre-training, fine-tuning, evaluation, and inference. Using Dynamic Workflows, the M3 Agent autonomously completed the entire task in 90 minutes. The test showed that after about 1000 training steps, the model evolved from generating random responses to demonstrating basic logic and Q&A capabilities, validating M3’s potential as an automated research tool.
Case Study 2: Refactoring and Optimizing an Existing Software Project
The second test focused on developing version 2.0 of a project called “Humanize PPT.” The project’s goal is to use AI to generate outlines and speaker modes for HTML presentations, requiring compatibility with existing frameworks like guizang-ppt-skill and frontend-slides. When previously using the GPT-5.5 model, developers encountered issues with the model “being lazy” and providing only temporary workarounds.
After switching to MiniMax M3, the agent refactored the code and resolved legacy issues within one hour. Notably, it took only 2 minutes and 50 seconds to read and analyze the entire project’s codebase and provide modification suggestions down to the specific line of code. Ultimately, M3 successfully implemented a feature to call sub-agents within the conversation to batch-generate all compatible themes at once, demonstrating its stability and reliability in handling complex code dependencies and executing long-term development plans.
Pricing and Market Positioning
With the launch of M3, MiniMax has adjusted the pricing model for its API services, moving from time-based credit refreshes to monthly subscription plans with a fixed number of tokens.
- Plus Plan: Offers 600 million tokens, priced at 49/month.
- Max Plan: Offers 1.8 billion tokens, priced at 119/month.
- Ultra Plan: Offers 5.5 billion tokens, priced at 469/month.
Considering its performance, efficiency, and pricing, MiniMax M3 is positioned as a productivity tool for developers and professional users, especially for scenarios that require building and running complex, long-running AI Agent applications.
