MiniMax M2.5 / M2.5-Lightning

// Overview

MiniMax M2.5 is an open-weight large language model from Shanghai-based MiniMax, released on February 12, 2026. It uses a sparse Mixture-of-Experts design with roughly 229B total parameters but only about 10B active per token, which keeps inference cheap while preserving frontier-level quality on coding, tool use, web search, and office-document tasks. The model is text-only; MiniMax handles vision and audio through separate models.

The release shipped in two variants that share the same weights and capability but differ in speed: the standard MiniMax M2.5 and a faster M2.5-Lightning. MiniMax positioned both as built for real-world productivity, trained with reinforcement learning across many complex digital working environments rather than tuned purely for static benchmarks. It posts 80.2% on SWE-bench Verified and 76.3% on BrowseComp, and lands around 42 on the Artificial Analysis Intelligence Index — competitive with much larger closed models.

The headline pitch is price. At launch MiniMax priced M2.5 at roughly one-tenth to one-twentieth the cost of comparable closed frontier models such as Claude Opus 4.6, making sustained agentic use economically viable. The weights are published on Hugging Face under a modified MIT license that requires commercial users to display the 'MiniMax M2.5' name in their interface. It has since been succeeded in MiniMax's lineup by later M-series models (M2.7), and on the official API M2.5 is now listed among legacy models.

// Benchmarks

1. SWE-bench Verified80.2%
2. Multi-SWE-Bench51.3%
3. BrowseComp76.3%
4. GPQA Diamond85.2%
5. AIME 202586.3%
6. MMLU-Pro80.1%
7. Artificial Analysis Intelligence Index42%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Launch pricing from MiniMax. Both variants support prompt caching. M2.5 is now listed among legacy models on the MiniMax API; third-party providers (OpenRouter, SiliconFlow, Fireworks, Novita) may price it differently.

Pricing source ↗

// Strengths

• Strong agentic coding: 80.2% on SWE-bench Verified, competitive with closed frontier models
• Very low cost per token relative to closed rivals, enabling long-running agent workloads
• Open weights under a permissive (modified MIT) license, downloadable and self-hostable
• Efficient MoE design — about 10B active parameters keeps serving cheap and fast
• Two serving tiers: standard M2.5 and a higher-throughput M2.5-Lightning variant
• Long 205K-token context window for large codebases and multi-step agent traces
• Built for real-world office and tool-use tasks, not just static benchmarks

// Best for

• Autonomous and semi-autonomous coding agents running across large codebases
• Cost-sensitive, long-horizon agent workflows where token volume is high
• Web-research and browsing agents (strong BrowseComp performance)
• Self-hosted or private deployment where open weights are a requirement
• Office and document-generation tasks (Word, Excel, PowerPoint workflows)
• Multilingual programming across Python, Go, C++, Rust, and TypeScript

// How to access

// MiniMax M-Series — every version

The full lineage of the MiniMax M-Series line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.