MiniMax M2.1

On January 22, 2026, MiniMax published a follow-up post on minimax.io titled "MiniMax M2.1: Post-Training Experience and Insights for Agent Models," explicitly naming MiniMax M2.1 as the subject and describing it as the latest flagship open-source model built on further post-training optimization over the M2 generation The same post breaks down M2.1's agentic data synthesis into three categories: real-data-driven synthesis called SWE Scaling, expert-driven synthesis for AppDev, and virtual long-horizon task synthesis called WebExplorer. SWE Scaling is described as leveraging GitHub pull requests and commits as a structured data sourc

Kilo Code's product page at kilo.ai/models/minimax-minimax-m2-1 explicitly names MiniMax-M2.1 as a lightweight, state-of-the-art large language model optimized for coding, agentic workflows, and modern application development, noting that with only 10 billion activated parameters it delivers a major jump in real-world The same page reports Kilo Code's PinchBench (OpenClaw) results for MiniMax-M2.1: an average score of 82.7% across 17 of 50 official models, average time of 19m 51s across 33 runs per OpenClaw task, and average cost of $0.168 per benchmark run. Category breakdown shows the best verified PinchBench v2 run scoring 95.1%