Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (IT Home · Nov 5, 2026)

Moonshot AI opens weights for 2.8T-parameter Kimi K3.1 Pro

KEY FACTS

  • Moonshot AI released the Kimi K3.1 Pro model weights on Hugging Face on November 5.
  • The model ranks first on the Agent Arena open-weights leaderboard and 13th across all models.
  • The ranking is based on over 17,500 real agent sessions, with a net improvement rate of +5.1%.
  • The model has 2.8 trillion total parameters, roughly 104 billion activated at inference, and supports a 1 million token context.
  • The BF16 weights are about 5.6TB, and even after INT4 quantization they still require about 1.4TB.

KEY DATA

2.8 trillionTotal parameters
~104 billionActivated inference parameters
~1 million tokensContext window
65.7%Terminal-Bench 4.0 score

PHYSIX OBSERVATION

Open weights do not mean everyone can run it: at 5.6TB, the size shuts most developers out, and the real beneficiaries are institutions with compute clusters. But Kimi took first place among open models on Agent Arena using real-session signals, showing that agent reliability is becoming the new competitive frontier, and closed-source models' price moat will be gradually eroded.

Source: IT Home report