MiniMax Releases MiniMax M2: A Mini Open Model Built for Max Coding and Agentic Workflows at 8% Claude Sonnet Price and ~2x Faster

Can open source MOE really support a proxy coding workflow at a fraction of the costs of the main model while maintaining long-term use of the tool across MCP, shell, browser, retrieval, and code? The MiniMax team has just launched Mini Max-M2a combination of the expert MoE model optimized for encryption and proxy workflow. The weights on Hugging Face are published under an MIT license, and the model is modeled for end-to-end use of the tool, multi-file editing, and long-term planning. It lists a total of 229 billion parameters with about 10 billion active per token, keeping memory and response time in check during proxy loops.

Structure and why does activation size matter?

Mini Max-M2 It is a consolidated MOE that addresses approximately 10 billion active parameters per token. Smaller activations reduce memory pressure and latency in the plan, act, and check loops, and allow more concurrent operation in the CI, browse, and retrieval threads. This is a performance budget that enables speed and cost claims compared to dense models of similar quality.

Mini Max-M2 It is a model of interwoven thinking. The research team encapsulated the internal logic ... blocks, and directs users to keep these blocks in their conversation history across turns. Removing these parts hurts quality in multi-step jobs and toolchains. This requirement is clear on the HF model page.

Standards targeting codecs and proxies

The MiniMax team reported a set of proxy and code evaluations closer to developer workflows than static QA. On the station bench, Table 46.3 is shown. On the Multi SWE Bench, it shows 36.2. In BrowseComp it shows 44.0. SWE Bench Verified is listed at 69.4 with scaffold details, and OpenHands with 128k100 context.

MiniMax’s official announcement confirms 8% off Claude Sonnet prices, nearly 2x speed, plus a free entry window. The same note provides the prices of the specific symbols and the trial deadline.

Compare M1 vs M2

Key takeaways

1. The M2 is shipped as open weights on Hugging Face under MIT, with safeties in F32, BF16 and FP8 F8_E4M3.
2. The model is an embedded MOE with a total of 229 billion parameters and about 10 billion actives per token, which the card associates with lower memory usage and more consistent latency in the planning, acting, and checking loops typical of agents.
3. Outputs encapsulate internal thinking ... The sample ticket explicitly directs keeping these clips in your chat history, warning that removal results in decreased multi-step performance and tool usage.
4. Reported results cover Terminal-Bench, (Multi-)SWE-Bench, BrowseComp, and others, with key observations for reproducibility, and day 0 service for SGLang and vLLM is documented with concrete deployment evidence.

Editorial notes

Available in open weights under MIT, MiniMax M2 is a combination of design experts with a total of 229 bytes of parameters and about 10 bytes per token, which targets proxy loops and coding tasks with less memory and more stable latency. Hugging Face ships in security tools in FP32, BF16, and FP8 formats, and provides deployment notes as well as a chat template. The API documents humanized endpoints and lists prices with a limited free window for evaluation. vLLM and SGLang recipes are available for local provisioning and benchmarking. Overall, the MiniMax M2 is a very solid unlocked version.

verify API doc, weights and repo. Feel free to check out our website GitHub page for tutorials, codes, and notebooks. Also, feel free to follow us on twitter Don’t forget to join us 100k+ mil SubReddit And subscribe to Our newsletter. I am waiting! Are you on telegram? Now you can join us on Telegram too.

Michel Sutter is a data science specialist and holds a Master’s degree in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michelle excels at transforming complex data sets into actionable insights.