February 13, 2026 — MiniMax formally launched its new‑era textual content mannequin, MiniMax M2.5, positioning it as a “manufacturing‑grade native agent mannequin”.
With solely 10 billion activated parameters, M2.5 delivers flagship‑degree efficiency. On SWE-Bench Verified, it scored 80.2%; on Multi-SWE-Bench, it ranked first — surpassing Claude Opus 4.6 in multi‑language, advanced environments. The mannequin reveals “native spec conduct,” proactively deconstructing structure and useful planning earlier than coding, carefully resembling a human architect.
M2.5 is constructed for full‑stack growth (frontend, backend, database) and Vibe Coding — turning pure language into executable system designs. It achieves 100+ tokens per second inference throughput (2x mainstream flagship fashions). At 100 tokens/sec, steady operation prices $1/hour; at 50 tokens/sec, $0.3/hour.
MiniMax attributes M2.5’s fast iteration — from M2 to M2.5 in 108 days, SWE-Bench rating leaping from 69.4 to 80.2 — to massive‑scale Agent Reinforcement Studying (RL Scaling). Its proprietary Forge framework decouples coaching engine from agent, enabling generalized optimization throughout any scaffolding, with 40x coaching acceleration by way of asynchronous scheduling and tree‑primarily based merging.
Internally, M2.5 now autonomously handles 30% of actual enterprise duties at MiniMax, overlaying R&D, product, gross sales, HR, and finance. In coding, M2.5‑generated code accounts for 80% of newly dedicated code.
M2.5 went stay on MiniMax Agent on February 12 and was globally open‑sourced the subsequent day. Inside 24 hours, customers constructed over 10,000 specialists on the platform.
Supply: QbitAI
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the newest breakthroughs, get unique updates, and join with a world community of future-focused thinkers.
Unlock tomorrow’s tendencies immediately: learn extra, subscribe to our publication, and develop into a part of the NextTech group at NextTech-news.com
