# Moonshot AI Kimi model development

> Live situation record from CLSTR: https://clstr.news/situations/moonshot-ai-kimi-model-development
> Updated: 2026-10-23T04:46:46.000Z. Sources: 36. Developments: 2.

Moonshot AI has expanded its Kimi model series, beginning with the release of Kimi K2.7-Code. This open-source, coding-focused agentic model utilizes a Mixture-of-Experts structure with 1 trillion total parameters and a 256K context window. It was designed to serve as a localized alternative for users facing connectivity and regulatory barriers when accessing international platforms.

Following this release, the company teased the upcoming Kimi K3.1 model via a preview video. The teaser demonstrated the Kimi Code programming agent’s ability to generate complex 3D scenes. Technical indicators suggest the K3.1 update may include expanded reasoning intensity settings and enhanced multi-agent collaboration capabilities.

On October 23, 2026, Moonshot AI released Kimi K3.1 Pro, a flagship model featuring a 3 trillion parameter mixture-of-experts architecture and a 1 million token context window. Designed for advanced agentic tasks such as coding, 3D modeling, and financial analysis, the model achieved high benchmark scores, including 65.7% on Terminal-Bench 4.0 and 92.7% on the BenchCAD 3D reconstruction test. In financial workflows, it reached a score of 64.4 on the OfficeQA Pro v2 benchmark. The model utilizes a mixture-of-experts (MoE) architecture that activates 104 billion parameters per task to optimize efficiency.

The company reported that internal testing showed factual error rates dropping from 23.4% in the previous K3 model to 8.3%. To mitigate security risks, Moonshot AI removed specific cyber-related data from training sets ahead of a planned open-weights release expected by November 4. A high-speed version offering two to three times faster output is also available via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

## Claims

- Moonshot AI released the Kimi K3.1 Pro model. (corroborated by 26 sources)
- The model has 3 trillion total parameters and uses a mixture-of-experts architecture. (corroborated by 22 sources)
- The model features a 1 million token context window. (corroborated by 20 sources)
- Moonshot AI plans to release the model weights by November 4, 2026. (corroborated by 18 sources)
- Kimi K3.1 Pro scored 65.7% on the Terminal-Bench 4.0 benchmark. (corroborated by 16 sources)
- A Highspeed version of the model is two to three times faster. (corroborated by 12 sources)
- The model scored 92.7% on the BenchCAD 3D modeling benchmark. (corroborated by 8 sources)
- Internal testing showed hallucination rates dropped from 23.4% to 8.3%. (corroborated by 5 sources)
- Moonshot AI removed cyber-related data from training to reduce exploit capabilities. (corroborated by 4 sources)

## Timeline

### 2026-10-23: Moonshot AI unveils Kimi K3.1 Pro with 3 trillion parameters

Moonshot AI released Kimi K3.1 Pro, a 3-trillion-parameter model with a 1-million-token context window, featuring improved coding, 3D modeling, and financial capabilities while limiting cyber-offensive data.

34 sources. https://clstr.news/cluster/moonshot-ai-teases-kimi-k31-model-with-preview-video

### 2026-10-04: Moonshot AI releases Kimi K2.7-Code open-source model

Moonshot AI released the open-source Kimi K2.7-Code model, showing strong performance in coding benchmarks. Meanwhile, Kimi Work is positioned as a domestic alternative to OpenAI's cloud-based Codex.

2 sources. https://clstr.news/cluster/moonshot-ai-releases-kimi-k27-code-open-source-model

---
Cite as: Moonshot AI Kimi model development. CLSTR, https://clstr.news/situations/moonshot-ai-kimi-model-development