MiniMax’s new open M2.5 and M2.5 Lightning near state-of-the-art while costing 1/20th of Claude Opus 4.6
Publish Date: 2026-02-13 02:51:00
Source Domain: venturebeat.com
-
Release and Open Source Model: Chinese AI startup MiniMax introduced its M2.5 language model, available in two variants and made open source on Hugging Face under a modified MIT license. It mandates attribution for commercial uses.
-
Cost Efficiency: The M2.5 model dramatically reduces the costs of high-end AI by up to 95%, making advanced AI more accessible and affordable for enterprise use by offering its API and partner APIs at very low rates.
-
Performance and Efficiency: The Mixture of Experts (MoE) architecture allows M2.5 to activate only 10 billion of its 230 billion parameters, enabling performance comparable to top-tier models at a fraction of the cost, especially in enterprise tool use.
-
Development and Training: MiniMax developed a proprietary Reinforcement Learning framework called Forge and used CISPO (Clipping Importance Sampling Policy Optimization) for stable training, achieving industry-leading results on various benchmarks.
-
Benchmark Achievements: M2.5 matches and even surpasses industry-leading models on key benchmarks, showcasing state-of-the-art performance in coding, search tool use, and multi-language coding.
-
Strategic Implications: The significant cost reduction and performance of M2.5 transforms operational strategies for enterprises, enabling widespread deployment of powerful AI for routine tasks, more efficient real-time applications, and large-scale automated code audits.
-
Affordable AI Agents: The affordability of M2.5 means developers can build “agents” that autonomously handle complex projects without excessive costs, positioning MiniMax as a leader in making AI more useful and widely applicable.
