Moonshot AI releases Kimi K2.5 with Agent Swarm that coordinates 100 parallel AI agents

Moonshot AI Unveils Kimi K2.5 as Most Powerful Open-Source Model

Moonshot AI, backed by Alibaba and HongShan (formerly Sequoia China), released Kimi K2.5 on Tuesday, positioning it as the world's most powerful open-source large language model to date 2

2

. Built on top of the Kimi K2 LLM that debuted last summer, this multimodal AI model was trained on 15 trillion mixed visual and text tokens, making it natively capable of understanding text, images, and video 1

1

. The model features a mixture-of-experts architecture with 1 trillion parameters, though only 32 billion parameters are activated at any given time, significantly reducing hardware usage while maintaining performance 4

4

.

In released benchmarks, Kimi K2.5 outperforms proprietary models from OpenAI, Anthropic, and Google on several key metrics. The model achieved 50.2% on Humanity's Last Exam (HLE) benchmark with tools, surpassing GPT-5.2 and Claude Opus 4.5 3

3

. On coding benchmarks, it scored 76.8% on SWE-Bench Verified and outperformed Gemini 3 Pro, while beating both GPT-5.2 and Gemini 3 Pro on the SWE-Bench Multilingual benchmark 1

1

. For video understanding, it surpassed GPT-5.2 and Claude Opus 4.5 on VideoMMMU, a benchmark measuring how models reason over videos.

Coding With Vision Generates Web Interfaces From Visual Inputs

What sets Kimi K2.5 apart is its ability to generate web interfaces based solely on images or video, a capability Moonshot calls "coding with vision" 2

2

. The model can recreate websites complete with interactive elements and scroll effects by analyzing recorded videos of existing sites from a user's screen perspective. While users can feed it text for traditional coding tasks, the model excels at understanding visual inputs and translating them into functional code 1

1

.

This marks a meaningful advancement for "vibe coding" tools, which rely on intuitive methods easily deployed by non-experts rather than traditional coding approaches. While ChatGPT, Claude, and Gemini can generate raw code from screenshots, Kimi K2.5 cuts out the intermediary step by directly producing finished, usable products 2

2

. "By reasoning over images and video, K2.5 improves image/video-to-code generation and visual debugging, lowering the barrier for users to express intent visually," Moonshot stated.

To enable developers to use these multimodal capabilities, Moonshot launched Kimi Code, an open-source coding tool rivaling Anthropic's Claude Code and Google's Gemini CLI 1

1

. Developers can access Kimi Code through terminals or integrate it with development software such as VSCode, Cursor, and Zed, with support for images and videos as input.

Agent Swarm Coordinates 100 Sub-Agents for Parallel Task Execution

A key differentiator for Kimi K2.5 centers on its Agent Swarm capability, which orchestrates up to 100 specialized sub-agents working in parallel 3

3

. The company compared this orchestration to a beehive where each agent performs a specific task while contributing to a common goal. The model learns to self-direct these sub-agents and can execute parallel workflows of up to 1,500 tool calls 5

5

.

By running multiple tasks concurrently, Agent Swarm significantly reduces end-to-end latency compared to sequential agent execution. Internal evaluations showed that end-to-end runtime could be reduced by up to 80% 2

2

. "Tasks that required days of work now can be accomplished in minutes," Moonshot stated, emphasizing that the real metric they care about is "how much of your day did AI actually give back to you" 3

3

.

Users with active Allegretto or Vivace Moonshot accounts, costing $31 per month and $159 per month respectively, can test Agent Swarm on the Kimi website by selecting "K2.5 Agent Swarm (Beta)" from the model drop-down menu 2

2

.

Growing Competition in AI Coding Assistants Market

The release comes as AI coding assistants become major revenue drivers for AI labs. Anthropic announced in November that Claude Code had reached $1 billion in annualized recurring revenue (ARR), adding $100 million to that figure by the end of 2025 according to Wired 1

1

. Moonshot's Chinese competitor, DeepSeek, is set to release a new model with strong coding capabilities next month, according to The Information.

Moonshot AI saw a 170% increase in users between September and November for Kimi K2 and Kimi K2 Thinking, which was released in early November 3

3

. The startup, founded by former Google and Meta AI researcher Yang Zhilin, raised $1 billion in a Series B round at a $2.5 billion valuation. Bloomberg reported that the company picked up $500 million in funding last month at a $4.3 billion valuation and is already seeking to raise a new round at a $5 billion valuation 1

1

.

The model's code is available on Hugging Face, allowing developers to download and modify its weights for customized applications 4

4

. However, with 1 trillion parameters, the model requires significant GPU resources to operate optimally, reflecting its focus on handling demanding computational workloads. The model's open-source nature and modular design ensure it can be tailored to meet specific user needs, positioning it as a compelling option for enterprises considering orchestration strategies where agents communicate and pass off tasks automatically rather than following rigid frameworks.