An open source vision-language model unleashes minion agents that enable it to perform tasks more quickly and effectively.
What’s new: Moonshot AI released Kimi K2.5, an updated version of its Kimi K2 large language model that adds vision capabilities and the ability to spawn what the authors call subagents — parallel workflows that control their own separate models to execute tasks as AI research, fact checking, and web development — and assign tasks to them.
- Input/output: Text, image, video in (up to 256,000 tokens); text out (109.5 tokens per second)
- Architecture: MoonViT vision encoder (400 million parameters), mixture-of-experts transformer (1 trillion total parameters, 32 billion active per token)
- Performance: Tops all other open-weights models in the Artificial Analysis Intelligence Index
- Availability: Free web user interface, weights free to download for noncommercial and commercial uses with attribution under modified MIT license, API $0.60/$0.10/$3.00 per million input/cached/output tokens, coding assistant $15 to $200 per month
- Features: Tool calls, web search, optional reasoning mode, subagents
- Undisclosed: Training data, training methods
How it works: Moonshot disclosed little information about how it built Kimi-K2.5. Among the details it revealed:
- Kimi-K2.5 is based on Kimi-K2-Base, a text-only model that was released in July. The team added a vision encoder and further pretrained the base model on 15 trillion image and text tokens.
- Using reinforcement learning, the team trained Kimi-K2.5, given a prompt, to generate subagents that operate in parallel, assign tasks to them, and incorporate their output into its response. Kimi-K2.5 received rewards for instantiating subagents and solving problems correctly. For instance, prompted to identify the top three YouTube channels across 100 domains, Kimi-K2.5 learned to gather information on each domain, generate 100 domain-specific subagents to search YouTube, and put their findings into a spreadsheet.
Results: In the Artificial Analysis Intelligence Index, a weighted average of 10 benchmarks, Kimi K2.5 with thinking mode switched on outperformed all other open-weights models tested. In Moonshot’s tests:
- Kimi K2.5 in thinking mode outperformed all open-weights models tested on various measures of reasoning, vision, coding, and agentic behavior. It also outperformed proprietary models including GPT 5.2 set to xhigh, Claude 4.5 Opus set to extended thinking, and Gemini 3 Pro set to high thinking on some vision and agentic benchmarks.
- Across 17 benchmarks of image and video performance, Kimi K2.5 achieved the highest score on 9, outperforming GPT 5.2 set to xhigh, Claude 4.5 Opus set to extended thinking, and Gemini 3 Pro set to high thinking.
- Subagents enabled Kimi-K2.5 to perform between 3 and 4.5 times faster than it did without using subagents. Subagents boosted its performance on the agentic benchmarks BrowseComp and WideSearch by 18.4 percentage points and 6.3 percentage points, respectively.
Yes, but: Moonshot didn’t disclose the cost of processing and memory incurred by Kimi K2.5’s use of subagents, so the tradeoff between speed/performance and processing/memory requirements is not clear.
Behind the news: Kimi K2.5 arrives 7 months after Moonshot’s initial vision-language model, the much smaller, 16 billion-parameter Kimi-VL, which also used the MoonViT vision encoder.
Why it matters: Building an agentic workflow can improve a model’s performance on a particular task. Unlike predefined agentic workflows, Kimi K2.5 decides when a new subagent is necessary, what it should do, and when to delegate work to it. This automated agentic orchestration improves performance in tasks that are easy to perform in parallel.
We’re thinking: Kimi K2.5 shifts task execution from chain-of-thought reasoning to agentic teamwork. Instead of responding to prompts sequentially, it acts as a manager of separate workflows/models that execute different parts of the job in parallel.