MiniMax's new open M2.5 and M2.5 Lightning near state-of-the-art while costing 1/20th of Claude Opus 4.6
Summary
MiniMax, a Shanghai-based AI startup, has released its M2.5 language model in two variants, M2.5 and M2.5-Lightning, which offer near state-of-the-art performance at significantly reduced costs. The models utilize a Mixture of Experts (MoE) architecture, activating only 10 billion of their 230 billion parameters per word, and were trained using a proprietary Reinforcement Learning (RL) framework called Forge, incorporating a mathematical approach named CISPO. M2.5 achieved an 80.2% on SWE-Bench Verified, matching Claude Opus 4.6 speeds, and scored 76.3% on BrowseComp and 51.3% on Multi-SWE-Bench. The standard M2.5 costs $0.15 per 1M input tokens and $1.20 per 1M output tokens, while M2.5-Lightning costs $0.30 per 1M input tokens and $2.40 per 1M output tokens, making them 1/10th to 1/20th the cost of competing models like Claude Opus 4.6.
Key takeaway
For CTOs and Machine Learning Engineers evaluating large language models for enterprise deployment, MiniMax's M2.5 models fundamentally alter the cost-benefit analysis. You can now implement advanced agentic workflows and high-context reasoning for tasks like document creation and coding without the prohibitive token costs, enabling broader automation and faster end-to-end task completion. Consider integrating M2.5 to scale your AI initiatives and reduce operational expenses significantly.
Key insights
MiniMax's M2.5 models offer near state-of-the-art AI performance at a fraction of the cost, enabling widespread agentic applications.
Principles
- Cost-efficiency drives AI adoption.
- Sparse MoE architectures enhance performance.
- RL with real-world environments improves agentic capabilities.
Method
MiniMax developed a proprietary Reinforcement Learning (RL) framework called Forge, utilizing CISPO (Clipping Importance Sampling Policy Optimization) to train its Mixture of Experts (MoE) model, M2.5, over two months.
In practice
- Deploy high-context models for routine tasks.
- Build agentic pipelines for real-time applications.
- Automate code audits with cost-effective models.
Topics
- MiniMax M2.5
- Mixture of Experts
- Reinforcement Learning
- AI Agents
- AI Cost-Efficiency
Best for: CTO, Machine Learning Engineer, NLP Engineer, AI Engineer, Director of AI/ML, MLOps Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.