MiniMax 2.5 (M2.5) is the latest frontier model that is specialized for coding, agentic workflows, search, and real-world productivity. As a small Mixture of Experts (MoE) model, it delivers one of the best cost to intelligence ratios for these use cases, hitting 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp (with full context management). It's 37% faster than its predecessor on complex agentic tasks, matching the throughput of Claude Opus 4.6 while costing a fraction of the price. SambaNova is running it fastest in the world—over 300 tokens/second.
