MiniMax M2.5

• 🆕 Coding/agentic LLM from MiniMax, released February 2026.
• 🧠 Adds enhanced reasoning with architect-style "spec" planning before coding.
• 📏 Roughly 198K-token context window in this catalog deployment.
• 🎯 Vendor reports 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench.
• ⚡ Reported 37% faster task completion than M2.1, fewer tokens used.
• 🔧 Function calling, web search, full development-lifecycle workflows.
• 🌐 Trained via RL across 10+ programming languages, many real environments.
• 💬 A high-throughput M2.5-highspeed variant targets low-latency use.

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax M2.5 is a text model from MiniMax, released in February 2026 and positioned for coding, agentic tool use, search, and office workflows. According to MiniMax's model card, it was trained with reinforcement learning across hundreds of thousands of complex real-world environments using the Forge agent-native RL framework and CISPO algorithm, and it spans the full development lifecycle from system design to testing across Web, Android, iOS, Windows, and Mac. A notable behavior MiniMax describes is "spec-writing": the model decomposes and plans features, structure, and UI before writing code.

Compared with its same-family predecessor M2.1, MiniMax reports meaningful efficiency gains: the company says M2.5 completes the SWE-Bench Verified evaluation about 37% faster while consuming fewer tokens per task (3.52M versus 3.72M), through improved task decomposition and more efficient chain-of-thought reasoning. On the provider's reported figures, M2.5 scores 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp with context management.

The model is offered with a standard build and an M2.5-highspeed variant tuned for higher throughput and lower latency, and a Mixture-of-Experts design keeps the active parameter set small relative to total size.

Within MiniMax's lineup, M2.5 was followed by later text releases MiniMax M2.7 and MiniMax M3, alongside the company's separate speech and music models.