Kimi K3.1 Pro technical announcement and cyber benchmark ablation

Evidence record

Moonshot vendor methodology and open-weight schedule; controlled xdev-bench-hard cyber risk-reduction ablation.

Read the original source ↗

Claim-level citations (4)

• supportsKimi K3.1 Pro: training-data ablation cuts a reported exploit benchmark from 90.5% to 81.0%: Moonshot reports xdev-bench-hard pass@3 scores of 90.5% for its unrestricted K3.1 variant, 81.0% for the risk-reduced K3.1 Pro and 33.3% for predecessor K3.

  Cyber risk-reduction training section and xdev-bench-hard figure

• supportsKimi K3.1 Pro: training-data ablation cuts a reported exploit benchmark from 90.5% to 81.0%: Moonshot reports xdev-bench-hard pass@3 of 90.5% for an unrestricted checkpoint, 81.0% for the cyber-risk-reduced K3.1 Pro, and 33.3% for predecessor K3.

  Cybersecurity risk mitigation; ablation results

• supportsKimi K3.1 Pro: training-data ablation cuts a reported exploit benchmark from 90.5% to 81.0%: The intervention removed selected offensive-security training and RL tasks; it did not remove all cyber ability and is not an independent real-world evaluation.

  Cybersecurity risk mitigation; evaluation design

• supportsKimi K3.1 Pro: training-data ablation cuts a reported exploit benchmark from 90.5% to 81.0%: Moonshot scheduled public full model weights for November 4, 2026; they were not released at the October announcement.

  Introducing Kimi K3.1 Pro; release statement