Sulat.com
AI models
Get 10-25% off from MiniMax
MiniMax Token Plan (minimax.io) logo

Model details

MiniMax-M2.5

MiniMax-M2.5 is designed as a productivity-oriented large language model that extends its predecessor's coding strengths into broader office tasks such as generating and operating Word, Excel, and PowerPoint files. A reviewer described it as scoring within roughly 0.6% of a leading frontier reasoning model on SWE-Bench Verified while operating at about one-twentieth of the cost, a claim that, if independently reproduced, would explain the strong interest around the release. The intended use therefore spans software engineering assistance, agentic coding, and document-oriented workflows where the model can both produce and manipulate real-world artifacts.

The architecture follows a Mixture-of-Experts design with around 230 billion total parameters but only about 10 billion active per inference, which is the central reason it can deliver frontier-tier capability without frontier-tier compute. It ships in Standard and Lightning variants offering different throughput profiles, and its open-weight release under a modified MIT license lets teams self-host, fine-tune, or run it through managed APIs. MiniMax trained it with a proprietary reinforcement learning framework called Forge that deployed the model across more than 200,000 real-world environments to ground the behavior in practical tasks rather than purely synthetic benchmarks, which suits its positioning for real-world productivity and coding pipelines.

MiniMax Token Plan (minimax.io)MiniMax-M2.5minimax

Quick Info

Powered by
Provider
MiniMax Token Plan (minimax.io)
Model key
MiniMax-M2.5
Release date
Feb 12, 2026
Last updated
Feb 12, 2026
Input modalities
Output modalities
Capabilities

Cost

A provider subscription or plan supersedes token-based pricing for this model.

Limits

Output tokens
131,072 tokens
Context window
204,800 tokens

Latest news about MiniMax-M2.5

MiniMax Token Plan (minimax.io)

Official sourceAnnouncement

MiniMax officially announced MiniMax M3 on June 1, 2026, positioning it as a frontier open-weight model for coding and agentic work. M3 introduces MSA (MiniMax Sparse Attention), a new attention architecture designed to sidestep the quadratic complexity of full attention and unlock a 1M-token context window, while nati Access is immediate for developers: M3 is available today through MiniMax Code, the Token Plan, and MiniMax's API services, with the Token Plan lineup on minimax.io listing M3 alongside M2.7 and M2.5 as supported models. For teams currently running MiniMax-M2.5 on the Token Plan, this signals that M2.5 is now a mid-tie

MiniMax Coding Plan (minimax.io)

Official sourceAnnouncement

MiniMax M2.5: Built for Real-World Productivity.

MiniMax Token Plan (minimax.io)

CoverageBenchmark

Analyze MiniMax-M2.5 API latency, throughput, and cost efficiency benchmarks. Compare response speed, token performance, and pricing for scalable AI applications.

MiniMax Token Plan (minimax.io)

CoverageBenchmark

OpenRouter lists MiniMax-M2.5 as a 205K-context model released Feb 12, 2026, described as a state-of-the-art LLM trained on complex real-world digital working environments that extends M2.1's coding expertise into general office tasks such as generating and operating Word, Excel, and PowerPoint files. Stated benchmarks Ten providers route the model on OpenRouter (Inceptron, DigitalOcean, Venice, StreamLake, AtlasCloud, Friendli, MiniMax, NovitaAI, SiliconFlow, Nebius, plus a MiniMax Highspeed tier), with P50 latencies ranging from 0.33s to 1.21s and throughputs of 25–75 tps, and uptime figures between 99.61% and 100%. A weighted-aver

MiniMax Token Plan (minimax.io)

Coverage

A Feb 13, 2026 Latent Space daily roundup reports that MiniMax released M2.5 during "China open model week," claiming an "Opus-matching" 80.2% score on SWE-Bench Verified. The newsletter frames M2.5 as a coding-focused model extending the capabilities of its predecessor M2.1, positioning it alongside the same week's ot The roundup is paywalled and bundles M2.5 into a multi-story headline alongside Gemini, Anthropic, and OpenAI news, so dedicated technical depth on M2.5 is shallow. No link to a first-party MiniMax release or model card is provided, and the benchmark figure is presented as a vendor claim. Still, it is a timely Feb 2026

Videos about MiniMax-M2.5

More models around MiniMax-M2.5