TECH

Moonshot AI Gets More Into Agents With New Model

Published: Jan. 28, 2026 5:48 p.m. GMT+8

00:00

00:00/00:00

Listen to this article 1x

Photo: Visual China

Chinese artificial intelligence (AI) startup Moonshot AI has released an upgraded version of its flagship model Kimi, which it hailed as a multimodal agent model good at handling tasks from coding to essay writing.

The open-source model, known as K2.5, can process text, images and videos from a single prompt and handle agentic workflows, where agents make decisions, take actions and coordinate tasks with minimal human intervention, Moonshot AI said on Tuesday.

You've accessed an article available only to subscribers

Subscribe today for just $.99.

VIEW OPTIONS

Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.

Subscribe to both Caixin Global and The Wall Street Journal — for the price of one.

Share this article

Open WeChat and scan the QR code

DIGEST HUB

Digest Hub

Back

Explore the story in 30 seconds

Moonshot AI released Kimi K2.5, a multimodal open-source AI model handling text, images, and videos, excelling in coding, essay writing, and agentic workflows.
K2.5 matches or surpasses major models: equaling OpenAI’s GPT-5.2 in coding, outperforming OpenAI and Anthropic in agentic tasks, and rivaling Google’s Gemini 3 Pro in image/video understanding.
China’s leading tech firms, including ByteDance, Alibaba, and Tencent, are rapidly building AI agent ecosystems; K2.5 is widely adopted by global firms.

AI generated, for reference only

Who’s Who

Moonshot AI: Moonshot AI, a Chinese AI startup, has released K2.5, an upgraded multimodal agent model. This model excels in various tasks, from coding to essay writing, processing text, images, and videos from a single prompt. K2.5 demonstrates competitive performance against models like OpenAI's GPT-5.2 and Google's Gemini 3 Pro in benchmarks, particularly in coding and agentic tasks. Moonshot AI shifted focus to coding and AI agents in 2025, with its K2 model gaining popularity.

OpenAI: In the provided text, OpenAI is mentioned as a prominent AI company. Its GPT-5.2 model is referenced, with Moonshot AI's K2.5 scoring similarly in coding benchmarks. OpenAI's "reasoning o3 and o4-mini models" are also noted for integrating various tools under ChatGPT.

Anthropic: Anthropic is one of the AI companies whose models are being benchmarked against Moonshot AI's new K2.5 model. Specifically, in the benchmark for handling agentic tasks, K2.5 has demonstrated superior performance compared to the latest models from Anthropic. In May, Anthropic launched its Claude 4 model, which significantly advanced software development with its advanced coding agents.

Google: Google's Gemini 3 Pro model is mentioned in the article as performing as well as Moonshot AI's K2.5 in image and video understanding. Google also repositioned Gemini as an AI assistant with the release of its 2.5 Pro version in May.

Manus: Manus is an AI startup that, in March, launched what it claimed was the first general-purpose AI agent. This agent is capable of performing diverse tasks, from resume screening to stock analysis.

ByteDance Ltd.: **ByteDance Ltd.** ByteDance, a Chinese internet technology company headquartered in Beijing, released its Doubao 1.6 model in June through its cloud unit, Volcano Engine. This release included 12 development tools specifically designed for AI agents, indicating ByteDance's expansion into the AI agent ecosystem.

Volcano Engine: Volcano Engine is ByteDance Ltd.'s cloud unit. In June, it released the Doubao 1.6 model, accompanied by 12 development tools specifically designed for AI agents. This initiative by Volcano Engine highlights the growing interest and investment in AI agent ecosystems among major tech companies in China.

Alibaba Group Holding Ltd.: Alibaba Group Holding Ltd. has built its own AI agent ecosystem. This indicates their involvement in the development and integration of AI agents within their various platforms and services, aligning with the broader trend among major tech companies to leverage advanced AI capabilities.

Tencent Holdings Ltd.: Tencent Holdings Ltd. has established its own AI agent ecosystem. The company recently integrated the third-party model DeepSeek into WeChat and hired former OpenAI researcher Yao Shunyu to lead its AI team. This highlights Tencent's investment and strategic moves in the evolving AI landscape.

Perplexity: The article mentions Perplexity as an AI search engine that has adopted Moonshot AI's K2 model. K2 is a trillion-parameter open-source model utilizing a Mixture of Experts (MoE) architecture, making it popular in both China and the U.S.

Genspark: Genspark is an AI firm that has adopted Moonshot AI's K2, a trillion-parameter open-source model. The K2 model utilizes a Mixture of Experts (MoE) architecture and is popular in both China and the U.S.

Vercel: Vercel is a cloud platform that has adopted Moonshot AI's K2 model. K2 is a trillion-parameter open-source model utilizing a Mixture of Experts (MoE) architecture. It is popular in both China and the U.S. and has also been adopted by other major players like Perplexity, Genspark, and YouWare.

YouWare: YouWare is a coding tool that has adopted Moonshot AI's K2 model. K2 is a trillion-parameter open-source model utilizing a Mixture of Experts (MoE) architecture, popular in both China and the U.S.

DeepSeek: DeepSeek is a third-party model that Tencent recently integrated into WeChat. Additionally, Tencent has recruited former OpenAI researcher Yao Shunyu to lead its AI team, which utilizes models like DeepSeek.

AI generated, for reference only

What Happened When