← All models

MiniMax-M2.5

Published Publisher MiniMaxMiniMax-M2.5

Overview

MiniMax-M2.5 is an open-weight language model designed for coding, tool-using agents, search and office productivity. MiniMax announced it on February 12, 2026. (minimax.io)

Capabilities and uses

  • Plans and develops software across multiple programming languages, including full-stack applications; it can assist with implementation, code review and testing.
  • Supports agent workflows involving task decomposition, tool calls and search, as well as document, spreadsheet and other office tasks.
  • MiniMax reported scores of 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp, the latter using context management. These are publisher-reported results under the stated evaluation setups, not guarantees of real-world performance. (minimax.io)
  • Weights are available for local deployment under a modified MIT license; users should review the license terms. (huggingface.co)

Pricing

At release, MiniMax's international announcement priced the approximately 50 tokens/second standard M2.5 at half the rate of its Lightning version: USD $0.15 per million input tokens and USD $1.20 per million output tokens. Both versions support caching, but the announcement does not specify cache rates. These are historical announcement prices, not verified current rates. Source: MiniMax release announcement, checked November 16, 2026. (minimax.io)

The international pricing overview and mainland China pricing overview, checked November 16, 2026, did not list M2.5 rates. Current direct-platform pricing therefore remains unverified. (platform.minimax.io)

Limitations

Results can be incorrect and should be reviewed, particularly for code execution, research and consequential decisions. Search and tool use depend on the surrounding tools and agent setup; benchmark results depend on their evaluation configuration. Availability, context limits and deployment costs may vary by service or deployment. (huggingface.co)

Official sources