Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (IT Home · Sep 9, 2026)

ModelBest Open-Sources MiniCPM5-2B, Leading Edge AI Agents

KEY FACTS

  • ModelBest and OpenBMB have open-sourced the 2-billion parameter model MiniCPM5-2B.
  • The model ranks first in overall scores among open-source base models under 4B parameters on the AA leaderboard.
  • Its Agentic Index score is 20 points, significantly outperforming other models in its class.
  • The companies also released the Meshy reinforcement learning framework and the JustRL II strategy.

KEY DATA

23 ptsAA Overall Score
20 ptsAgentic Index
53.9 ptsBenchmark Avg. Score
891 ptsReal Task Score

PHYSIX OBSERVATION

Small-parameter models are evolving from simple chatbots into edge-side agents capable of tool invocation. MiniCPM5-2B has set a new benchmark for its class with a significant lead, proving that high-density training and specialized reinforcement learning frameworks can effectively overcome computational bottlenecks. This marks a substantial reduction in the technical barriers to deploying general-purpose intelligent agents on mobile devices and other edge hardware, accelerating the implementation of edge AI applications.

Source: IT Home report