Moonshot AI launches Kimi K3.1 Pro with 2.8T parameters and 1M context

Moonshot AI has released Kimi K3.1 Pro, the first model in its K3.1 series. Built on a multimodal mixture-of-experts architecture that routes each token through a subset of specialised neural sub-networks, the model is available on kimi.ai, the Kimi mobile apps for iOS, Android and HarmonyOS, Kimi Work v3.1.0+ for Windows and Apple silicon Macs, Kimi Code via the /model command and the Kimi API as kimi-k3.1-pro with Zero Data Retention. A companion variant, Kimi K3.1 Pro Highspeed (kimi-k3.1-pro-highspeed), runs at two to three times the throughput of the standard model. Moonshot AI has not announced updated API token pricing on its launch post.

On independent and internal evaluations, Kimi K3.1 Pro gains 11 points over Kimi K3 max on the Artificial Analysis Intelligence Index to rank fourth overall at 55, while placing first among openly available models on the Artificial Analysis Cyber Index at 64%. With tools enabled on Kimi Hallu Bench Hard, its claim-level hallucination rate drops to 8.3% from 23.4% on Kimi K3, compared with 5.5% for GPT-6 Astra and 15.0% for Claude Opus 5.5. To limit offensive misuse, Moonshot AI removed targeted cyber-exploitation datasets during mid-training and filtered reinforcement learning rollouts, reducing pass@3 on its internal xdev-bench-hard cyber benchmark from 90.5% to 81.0%.

Alongside the model, Moonshot AI introduced 3D Make inside kimi.ai and the Kimi app so users can preview, edit and export generated 3D assets directly for 3D printing. In a long-horizon biomechanics demonstration over the open-source human-atlas model (296 skeletal and 402 muscle components), Kimi K3.1 Pro ran 167 autonomous missions and 420 self-reviews, diagnosing 36 shoulder-muscle moment-arm calculation bugs to achieve 11.2 seconds of continuous bipedal walking.

Key numbers

• Parameters: 2.8T total parameters with 104B active per token
• Highspeed throughput: 2 to 3 times the generation speed of Kimi K3.1 Pro
• Cyber mitigations: 81.0% pass@3 on internal xdev-bench-hard, down from 90.5% before filtering
• Open-weight release: full model weights scheduled for release by 4 November 2026