MiniMax M2: A 230‑B Parameter MoE Model That Packs Power into 10 B
MiniMax M2 is a cutting‑edge 230 B‑parameter Mixture‑of‑Experts (MoE) LLM that delivers top‑tier performance in math, science, coding, and tool‑use while staying compact, fast, and cost‑effective. Developed and maintained by MiniMax (see MiniMax.io), the model’s active capacity is 10 B parameters, yet it retains the full power of a 230 B system. The MiniMax‑M2 model is released with no base model and is announced on GitHub (announcement). Its architecture excels in advanced reasoning and multi‑turn preservation , making it a versatile choice for a wide range of AI applications.
MiniMax M2: A New Benchmark for Open‑Source Large Language Models
MiniMax M2 introduces a suite of breakthrough innovations that set it apart from existing open‑source LLMs. Its 230 B total parameters are managed through an efficient Mixture‑of‑Experts (MoE) architecture that activates only 10 B at a time, delivering unprecedented performance while keeping the model compact, fast, and cost‑effective. Ranked #1 globally among open‑source models by Artificial Analysis, MiniMax M2 excels in mathematics, science, instruction following, coding, and agentic tool use. It offers advanced end‑to‑end coding capabilities —including multi‑file edits, coding‑run‑fix loops, and test‑validated repairs—alongside complex long‑horizon toolchain planning across shell, browser, retrieval, and code runners. The model’s interleaved thinking format enhances reasoning, and its turn‑to‑turn reasoning preservation ensures continuity in multi‑turn interactions. These features collectively yield lower latency, lower cost, and higher throughput for both interactive agents and batched sampling.
Key Innovations
- 230 B total parameters with only 10 B activated via an efficient MoE architecture
- #1 ranking among open‑source models (Artificial Analysis)
- Superior intelligence in math, science, instruction following, coding, and agentic tool use
- Compact, fast, and cost‑effective design optimized for deployment and scaling
- Advanced end‑to‑end coding : multi‑file edits, coding‑run‑fix loops, test‑validated repairs
- Complex long‑horizon toolchain planning across shell, browser, retrieval, and code runners
- Lower latency, lower cost, higher throughput for interactive agents and batched sampling
- Interleaved thinking model format for enhanced reasoning
- Reasoning preservation between turns for seamless multi‑turn interactions
Possible Applications for MiniMax M2
MiniMax M2’s compact yet powerful 230 B‑parameter MoE architecture, combined with its advanced reasoning and multi‑turn preservation, makes it especially possible for a handful of high‑impact use cases. Software development and coding assistance can benefit from its end‑to‑end coding capabilities, including multi‑file edits and test‑validated repairs. Agentic workflows and automation are enabled by its complex long‑horizon toolchain planning across shell, browser, and code runners, allowing the model to orchestrate multi‑step tasks autonomously. Finally, multi‑file editing and code repair can be performed efficiently thanks to the model’s fast, low‑latency inference and cost‑effective design. Each application must be thoroughly evaluated and tested before use.
- Software development and coding assistance
- Agentic workflows and automation
- Multi‑file editing and code repair
Common Limitations of Large Language Models
Large language models, despite their impressive capabilities, still exhibit several common limitations that users must be aware of. They can hallucinate or generate plausible but factually incorrect information, especially when dealing with niche or rapidly evolving topics. Their outputs are heavily influenced by the biases present in the training data, which can lead to unintended or harmful content. LLMs lack true world knowledge or grounding in real‑time data, making them unreliable for tasks that require up‑to‑date facts or precise factual verification. Additionally, they can struggle with long‑term context retention , often losing track of earlier conversation turns, and they may produce inconsistent or contradictory responses. Finally, the computational resources required for training and inference can be substantial, raising concerns about energy consumption and environmental impact.
MiniMax M2: A New Open‑Source LLM Milestone
MiniMax M2 marks a significant leap for open‑source large language models. With a 230 B‑parameter Mixture‑of‑Experts architecture that activates only 10 B at a time, it delivers world‑leading performance in mathematics, science, coding, and agentic tool use while remaining compact, fast, and cost‑effective. Ranked #1 globally among open‑source models by Artificial Analysis, it excels in advanced end‑to‑end coding (multi‑file edits, run‑fix loops, test‑validated repairs) and complex long‑horizon toolchain planning across shell, browser, retrieval, and code runners. Its interleaved thinking format and turn‑to‑turn reasoning preservation enable deeper, more consistent reasoning, and its lower latency and higher throughput make it ideal for interactive agents and batch inference. While MiniMax M2 opens exciting possibilities for software development, agentic automation, and multi‑file code repair, users should still thoroughly evaluate and test each application, keeping in mind the common limitations of LLMs such as hallucinations, bias, and context drift.