Kimi K3.1 Pro FAQ: Architecture, Pricing, and Open-Weight Plans

Basic Questions ​

1. When was Kimi K3.1 Pro released? ​

October 24, 2026. Moonshot AI unveiled the flagship model in the early morning of October 24 (Beijing time); CCTV Finance and Wall Street CN reported it the same day. It is available on the Kimi website, App, Kimi Work, Kimi Code, and Kimi API.

2. How is K3.1 Pro related to Kimi K3? ​

K3.1 Pro is a post-training upgrade of Kimi K3 and the first model of the K3 series. Moonshot AI says it is not trained from scratch; it substantially increases post-training compute on top of K3 and optimizes the reinforcement-learning framework, long-context reasoning, training efficiency, sandbox infrastructure, and task scheduling.

3. What are the total and active parameter counts? ​

2.8 trillion total (the vendor's phrasing: "3-trillion-class"), with about 104 billion active per token. The architecture is Mixture of Experts (MoE) built on Kimi Delta Attention (KDA) and Attention Residuals.

4. What is the context window? ​

1 million tokens. This matches the previous-generation K3 and supports long-horizon coding, knowledge work, and deep reasoning.

5. Which input modalities are supported? ​

Text, image, and video, with text output. The model has native vision understanding for engineering and creative tasks that use screenshots and visual feedback.

6. What is the API pricing? ​

Per million tokens: CNY 13 input (cache miss), CNY 1.3 input (cache hit), CNY 68 output. In USD (citing Artificial Analysis): $2 input, $10 output, $0.20 cache hit. Moonshot AI also released the Highspeed variant, priced at 2x the standard version.

7. What is the Highspeed variant? ​

An edition with roughly 2-3x the output speed and 2x the price of the standard version, with comparable intelligence. It was announced alongside the standard model on release day.

8. When will the model weights be open? ​

Full weights are planned before November 4, 2026. Actual usage scope is subject to the final license terms.

9. What security assessment was done before opening weights? ​

Moonshot AI identified and removed cyber attack-defense-related data from mid-training and reinforcement-learning stages before opening weights. On its in-house xdev-bench-hard test, the model without removal scored 90.5% and the final version 81.0%; the company says the data cleaning reduced vulnerability-research ability.

10. Where can I use it? ​

Kimi website, App, Kimi Work, Kimi Code, and Kimi API. Enterprises can also use Kimi Hosted Agents for cloud-hosted long-running agent workflows.

Advanced Questions ​

11. How does it compare with closed-source flagships? ​

It reaches or approaches top closed-source models on several benchmarks. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, above OpenAI GPT-6 Astra's 58.2% and below Anthropic Claude Opus 5.5's 66.4%; the Artificial Analysis Intelligence Index gives it 55, 4th among public models. Some scores come from internal benchmarks and are not directly comparable across evaluation systems.

12. Which capabilities were strengthened most? ​

Long-horizon coding, 3D modeling and spatial reasoning, finance research, and long-horizon agent execution. Official examples: BenchCAD (w/ tools) 92.7%, OfficeQA Pro v2 64.4 (vendor claims SoTA); internal financial modeling up ~14% and finance research up ~20% vs K3; on the Kimi Hallu Bench Hard stress test, the single-fact error rate dropped from 23.4% (K3) to 8.3%.

Data Sources and References ​

• Wall Street CN (2026-10-24): http://m.toutiao.com/group/7700018343629046308/
• CCTV Finance (2026-10-24): http://m.toutiao.com/group/7700021695326847522/
• Kimi official (K3 tech blog): https://www.kimi.ai/blog/kimi-k3
• Kimi API platform (K3 docs): https://platform.kimi.com/docs/guide/kimi-k3-quickstart
• AIToolsRecap (2026-10-24, citing Artificial Analysis): https://aitoolsrecap.com/Blog/ai-news-october-24-2026