Chinese company Moonshot AI has upgraded its open-source KM K2 model, transforming it into a coding and vision model with an architecture that supports agent swarm orchestration.
The new model, the Moonshot KM K2.5, is a good choice for enterprises that want agents that can take actions automatically rather than with a framework as the central decision-maker.
The company characterized the Kimi K2.5 as an “all-in-one model” that supports both visual and text input, allowing users to take advantage of the model for more visual coding projects.
Moonshot did not publicly disclose K2.5’s parameter count, but the KM K2 model on which it is based had a total of 1 trillion parameters and 32 billion active parameters due to its expert mixing architecture.
It’s the latest open-source model to provide an alternative to more closed options from Google, OpenAI and Anthropic, and it outperforms them on key metrics including agentic workflow, coding and vision.
But Humanity’s Last Test (HLE) Benchmark, KM K2.5 scored 50.2% (with tools), leaving behind OpenAI’s GPT-5.2 (xhigh) and Cloud Opus 4.5. also achieved this 76.8% But SWE-Bench VerifiedCementing its position as a top-tier coding model, however, GPT-5.2 and Opus 4.5 overtook it here at 80 and 80.9 respectively.
Moonshot said in a press release that it has seen a 170% increase in users of the Kmi K2 and Kmi K2 Thinking between September and November, which was released in early November.
Agent Swarm and built-in orchestration
Moonshot aims to take advantage of self-directed agents and the agent swarm paradigm built into KM K2.5. Agent swarms have been seen as the next frontier in enterprise AI development and agent-based systems. It has attracted a lot of attention in the last few months.
For enterprises, this means that if they build agent ecosystems with KM K2.5, they can expect to scale more efficiently. But instead of “scaling” or increasing model size to create larger agents, it is betting on creating more agents that can essentially organize themselves.
Kimi K2.5 “creates and coordinates a swarm of special agents working in parallel.” The company has compared it to a beehive where each agent performs a task, contributing to a common goal. The model learns to self-direct up to 100 sub-agents and can execute parallel workflows of up to 1,500 tool calls.
“Benchmarks only tell half the story. Moonshot AI believes that AGI should ultimately be evaluated by its ability to efficiently complete real-world tasks under real-world time constraints. The real metric they care about is this: How much of your day did the AI actually give back to you? Running in parallel substantially reduces the time required for a complex task – tasks that used to require days of work can now be completed in minutes. “Could,” the company said.
Enterprises considering their orchestration strategies have begun to look at agentic platforms, where agents communicate and pass tasks around rather than following a rigid orchestration framework that dictates when an action should be completed.
While KMI K2.5 may provide a compelling option for organizations that want to use this form of orchestration, some may feel more comfortable avoiding the agent-based orchestration involved in models and instead using a separate platform to separate model training from the agentic work.
This is because enterprises often want more flexibility in which models their agents create, so they can create an ecosystem of agents that tap the LLMs that work best for specific tasks.
Some agent platforms, such as Salesforce, AWS Bedrock, and IBM, provide separate observation, management, and monitoring tools that help users organize AI agents built with different models and enable them to work together.
Multimodal Coding and Visual Debugging
The model lets users code visual layouts, including user interfaces and interactions. It uses images and videos to understand the actions encoded in visual input. For example, K2.5 can reconstruct a website’s code by analyzing video recordings of a site, translating visual cues into interactive layouts and animations.
Moonshot said, “Interfaces, layouts, and interactions that are difficult to accurately describe in language can be communicated through screenshots or screen recordings, which the model can interpret and transform into fully functional websites. This vibe enables a new class of coding experiences.”
This capability is integrated into Km Code, a new terminal-based tool that works with VSCode and IDEs like Cursor.
it supports "autonomous visual debugging," Where the model inspects its own output – such as a rendered web page – references the documentation, and iterates over the code to fix layout changes or aesthetic errors without human intervention.
Unlike other multimodal models that can create and understand images, KMI K2.5 can create frontend interactions for websites with visuals, not only with the code behind them.
API pricing
Moonshot AI has priced the K2.5 API aggressively to compete with major US labs, offering significant price reductions compared to its previous K2 Turbo model.
- Input: 60 cents per million tokens (a 47.8% Reduce).
-
Cached Input: 10 cents per million tokens (a 33.3% Reduce).
-
Output: $3 per million tokens (a 62.5% Reduce).
The low cost of cached input ($0.10/M tokens) is particularly relevant "agent swarm" Features that often require maintaining large context windows across multiple sub-agents and extensive tool use.
Modified MIT License
While Kimi K2.5 is open-source, it is released under a modified MIT license that includes a specific segment targeting "hyperscale" Commercial users.
A license provides standard permissions to use, copy, modify, and sell the software.
However, it stipulates that if the software or any derivative work is used for a commercial product or service that has more than 100 million monthly active users (MAU) or monthly revenues more than $20 million USD, the entity must prominently display "km K2.5" On the user interface.
This clause ensures that while the model remains free and open to the vast majority of the developer community and startups, major tech giants cannot white-label Moonshot’s technology without providing visible attribution.
it is not full "open source" But it’s better than Meta’s similar llama licensing terms "open source" Family of models, which required those companies with 700 million or more monthly users to obtain a special enterprise license from the company.
What this means for modern enterprise AI builders
For practitioners defining the modern AI stack – from LLM decision-makers optimizing deployment cycles to AI orchestration leaders deploying agents and AI-powered automated business processes – KM K2.5 represents a fundamental shift in leverage.
By embedding swarm orchestration directly into models, Moonshot AI effectively hands these resource-constrained builders a synthetic workforce, allowing a single engineer to direct a hundred autonomous sub-agents as easily as a single signal.
it "To drag out" The architecture directly addresses data decision makers’ dilemma of balancing complex pipelines with limited headcount, while the low pricing structure transforms high-context data processing from a budget-breaking luxury into a regular commodity.
Ultimately, K2.5 suggests a future where the primary constraint on an engineering team is no longer the number of hands on keyboards, but the ability of its leaders to choreograph a herd.
<a href