News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnMiniMax-M2.5

MiniMax-M2.5

Cost-efficient coding and agentic model from MiniMax AI

MiniMax AI released MiniMax-M2.5, a language model optimised for coding, search, and office tasks. It uses reinforcement learning to achieve high speed and low cost, aiming to make complex agents economically viable.

Share on WhatsApp

  • Costs $1 to run continuously for one hour at 100 tokens per second
  • Scores 80.2% on SWE-Bench Verified coding benchmark
  • Completes tasks 37% faster than the previous M2.1 model
  • Trained in hundreds of thousands of real-world environments
  • Supports over 10 programming languages including Rust and Go

What it is

MiniMax-M2.5 is a large language model developed by MiniMax AI. Released in November 2026, it focuses on coding, autonomous tool use, and professional office work. The model is designed to be highly efficient, offering fast inference speeds and low operational costs. It aims to support complex software development and administrative tasks without the high expenses typical of frontier models.

How it came about

The M2 series saw rapid iteration between late October and November 2026. MiniMax AI released M2, followed by M2.1, and then M2.5 within three and a half months. The company states this pace exceeded their initial expectations. Each version improved performance on benchmarks like SWE-Bench Verified, outpacing the progress rates of competing model families from other major labs.

Inside the model

The model was trained using extensive reinforcement learning in hundreds of thousands of simulated real-world environments. MiniMax AI built an internal framework called Forge to manage this agent-native training. This system decouples the training engine from the agent, allowing for better generalisation across different tools. The training data includes over ten programming languages and inputs from professionals in finance, law, and social sciences to improve office task performance.

What it can do

MiniMax-M2.5 scores 80.2 per cent on SWE-Bench Verified and 76.3 per cent on BrowseComp. It completes coding evaluations thirty-seven per cent faster than its predecessor, M2.1. The model plans software architecture before writing code and handles full-stack development. In office scenarios, it achieves a fifty-nine per cent win rate against mainstream models. It also performs expert-level search tasks with fewer rounds than previous versions.

How to use it

The model is available via API in two versions: M2.5 and M2.5-Lightning. Both have identical capabilities but differ in speed. The Lightning version processes one hundred tokens per second, while the standard version handles fifty. Pricing is set at zero dollars and thirty cents per million input tokens for the faster version. Running the model continuously for one hour costs one dollar at maximum speed.

Questions people ask about MiniMax-M2.5

How much does it cost to use MiniMax-M2.5?

The M2.5-Lightning version costs $0.30 per million input tokens and $2.40 per million output tokens. Running the model continuously for one hour at 100 tokens per second costs just $1. This is significantly cheaper than many competing frontier models.

What programming languages does MiniMax-M2.5 support?

The model supports over ten languages, including Python, Java, JavaScript, TypeScript, C, C++, Go, Rust, Kotlin, PHP, Lua, Dart, and Ruby. It is trained to handle full-stack development across web, mobile, and desktop platforms.

How fast is MiniMax-M2.5 compared to other models?

The M2.5-Lightning version operates at 100 tokens per second, which is nearly twice the speed of other frontier models. It completes complex coding evaluations thirty-seven per cent faster than its predecessor, M2.1, matching the speed of Claude Opus 4.6.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MiniMax-M2.5 yet, so the facts on this page come from the sources listed above: huggingface.co/MiniMaxAI/MiniMax-M2.5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.