The all-in-one visual assistant is here! Kimi has quietly launched K2.5, which can simultaneously dispatch 100 agents, boosting efficiency by up to 4.5 times.

The all-in-one visual assistant is here! Kimi has quietly launched K2.5, which can simultaneously dispatch 100 agents, boosting efficiency by up to 4.5 times.

```

Moonshot has quietly made major upgrades to its flagship model, launching a new version called Kimi k2.5, marking key progress in multimodal capabilities and agent cluster collaboration. This upgrade was not accompanied by a public launch event, but was directly rolled out on the product side, aiming to consolidate its leading position in China’s highly competitive artificial intelligence market through substantial technical improvements.

According to a statement released by the company on Tuesday, the latest K2.5 model adopts a native multimodal architecture, capable of simultaneously processing text, images, and video through a single prompt. The model’s most notable technical breakthrough is its “agent cluster” capability, able to autonomously schedule up to 100 sub-agents to work in parallel. This architecture shortens the execution time of complex tasks by up to 4.5 times compared to a single-agent configuration, greatly improving processing efficiency.

This move comes during fierce competition among China’s top large model vendors. On the eve of a major update expected from rival DeepSeek, Moonshot has demonstrated its technical reserve with this upgrade. According to sources cited by Bloomberg, Moonshot recently completed a funding round, raising $500 million from investors including Alibaba and IDG Capital, bringing its post-money valuation to $4.3 billion. The company is now seeking a new round of funding at a valuation of up to $5 billion.

Currently, Kimi k2.5 is available to users through the Kimi.com web version, App, and API platform. This upgrade not only strengthens its traditional advantage in long-text domains but also introduces visual understanding and automatic code generation tools, aiming to achieve a more favorable position in enterprise applications and developer ecosystems.

Native Multimodality and Visual Reasoning

Kimi k2.5 achieves a leap in its basic architecture from single text processing to an all-around visual assistant. The model was pre-trained with approximately 15T of mixed visual and text tokens, so it no longer relies solely on simple OCR text recognition, but also possesses deep visual understanding capabilities.

According to the official description, users can directly upload complex circuit diagrams, handwritten math formulas, or financial statements, and K2.5 is able to understand the logic and carry out analytical reasoning. In programming, the model demonstrates powerful visual coding abilities, able to generate complete front-end interface code directly based on image or video input, and supports visual debugging, lowering the barrier for users to express development intent in visual ways.

Agent Cluster Collaboration Technology

The core highlight of this update is the “agent cluster” paradigm. K2.5 introduces Parallel Agent Reinforcement Learning (PARL) technology, allowing it to act as a coordinator to autonomously manage and orchestrate a cluster of 100 sub-agents.

When handling large-scale searches or complex workflows, the model can decompose tasks into parallelizable subtasks without pre-defining roles or workflows. For example, when screening video creators in a specific field, K2.5 can simultaneously launch 100 sub-agents for parallel searching and data aggregation. Internal test data shows that this parallel processing not only reduces end-to-end running time by 80% but also supports parallel workflows of up to 1,500 coordination steps, significantly breaking through the bottleneck of traditional single-agent capabilities.

Benchmarking and Office Efficiency

In terms of performance metrics, Moonshot claims K2.5 outperforms open-source peers across multiple benchmarks. The model stands out in agent benchmarks such as HLE, BrowseComp, and SWE-Verified. Particularly in programming and logical reasoning, K2.5 has narrowed the gap with top proprietary models, and its automated coding tool is designed to compete with Anthropic PBC’s Claude Code.

For actual office scenarios, K2.5 showcases the ability to handle high-density, large-scale knowledge work. Internal “AI Office Benchmarking” shows that compared to the previous K2 Thinking version, the new model’s performance in end-to-end tasks such as processing documents, spreadsheets, and building financial models has improved by 59.3%.

Funding Background and Market Landscape

Moonshot was founded by former Tsinghua University professor Yang Zhilin, who has AI project experience at both Meta Platforms Inc. and Google. Although the company is advancing commercialization through subscription plans and enterprise services, it still faces fierce competition for market share from rivals like Zhipu and MiniMax Group Inc., both of which recently raised more than $1 billion combined in Hong Kong IPOs.

With the DeepSeek R1 model’s breakthrough success in early 2025, China’s “Hundred Model War” in large model markets has entered the knockout stage, with many small players dropping out due to the pace of technology upgrades and capital demands. By releasing K2.5 ahead of competitors and matching it with a new high-valuation funding plan, Moonshot aims to prove its ongoing leadership in technological iteration and capital attraction.

Risk Warning and DisclaimerMarkets have risks, investment needs caution. This article does not constitute personal investment advice nor consider the specific investment objectives, financial situation, or needs of any individual user. Users should consider whether any opinions, views, or conclusions in this article are suitable for their situation. Investment based on this article is at your own risk.

```