Kimi K2.5: Visual Coding, Agent Swarms, Benchmarks and Local Deployment

Kimi K2.5 is an open-weight model from Moonshot AI that emphasizes coding, vision and agentic work. I am impressed by the release, particularly the beta Agent Swarm feature. The model can be accessed through Kimi’s services or downloaded for local deployment, subject to the resources needed to run it.

Moonshot describes a swarm that can coordinate up to 100 sub-agents and 1,500 tool calls, and reports up to a 4.5-fold reduction in execution time for its compared setup. Those numbers describe the company’s tests, not a guaranteed speedup on every task. I would like to try it with my own ClaudeBot, though I have not tested that integration yet.

According to Moonshot, K2.5 continues pretraining from Kimi K2-Base on approximately 15 trillion mixed visual and text tokens. It is natively multimodal and designed to handle visual inputs alongside language and tool use. In the published agentic comparisons, I see Kimi doing particularly well on BrowseComp and DeepSearchQA.

The displayed chart reports 74.9 on BrowseComp with context management, and 76.8 on SWE-bench Verified for coding. Kimi trails GPT-5.2 and Claude Opus 4.5 on that displayed coding comparison while exceeding the shown Gemini 3 Pro score. These are Moonshot’s reported results for specified model and test configurations, not a universal ranking.

The vision chart shows 78.5 on MMMU-Pro, behind the displayed GPT-5.2 and Gemini 3 results but above Claude Opus 4.5; the video tests also show strong results on the selected measures. I see different strengths across models, although one benchmark chart cannot establish that a model is best for all coding or visual work.

Cost is another part of the comparison. Moonshot’s chart places Kimi at a lower reported cost for several tasks, while Berman points to attractive benchmark performance. Price, test settings and the amount of compute used all affect comparisons, so the chart is a starting point rather than a fixed verdict.

The release page shows examples of websites with smooth motion and colorful designs. In one visual-coding demonstration, Kimi recreates a site from visual input without being given its source code. Moonshot attributes this to joint vision-and-text pretraining; the example is a selected demonstration, not evidence that every design will be reproduced exactly.

Another example asks the model to find the shortest route through a complicated maze image. It uses Python to process the image, implement breadth-first search and draw the route; breadth-first search is an algorithm, not a library. The displayed output labels the path as roughly 113,000 steps, and a closer view makes it easier to inspect.

A further Kimi Code demonstration translates the aesthetic of Matisse’s La Danse into an app. Berman describes an iterative visual-debugging loop: write code, inspect a rendered image, change the code and repeat. The shown sequence is sped up and illustrates the workflow rather than providing a controlled performance measurement.

Moonshot says its parallel agent reinforcement learning approach trains K2.5 to direct a swarm. An orchestrator breaks a complex job into parts, delegates them to specialized agents and combines their outputs. I compare that workflow with ClaudeBot and want to try it locally because I do not want to send my own ClaudeBot data to remote servers.

The orchestration diagram shows agents such as researchers, fact-checkers and web developers receiving assigned tasks. Moonshot reports up to an 80% reduction in end-to-end runtime in internal evaluations, with gains varying by workload. Parallel work can shorten the critical path, but coordination, dependencies and the slowest subtask still matter.

Berman reviews charts comparing agent performance and task complexity, then watches a demonstration assigning research across 100 YouTube subject areas. Each sub-agent produces an analysis and the coordinator assembles the responses. The visual shows how a broad task can be divided; it does not independently verify every result.

Moonshot also demonstrates knowledge-work tasks: annotating a Word document, constructing financial models with pivot tables, writing LaTeX equations in a PDF and creating a slide deck. Berman notes that some office-task comparisons use K2 Thinking as the baseline. A polished output still needs review for correctness.

The broader benchmark table compares K2.5 with several contemporaneous frontier models in reasoning, vision, coding and agentic search. Results vary by benchmark: the video highlights strong visual results and competitive coding, while some other models lead selected tests. For example, the vendor’s agentic search table distinguishes BrowseComp scores with and without context management and with Agent Swarm.

Benchmarks are useful as an initial signal, but Berman raises the possibility that models may be optimized for particular tests. I want to try the model on actual tasks, including building applications, rather than relying solely on the scorecard. Moonshot also makes an API available for people to run their own evaluations.

For a live test, I ask Kimi to compare its API prices with Claude Opus 4.5, GPT-5.2 and Gemini 3. It searches the web, writes Python and produces a comparison chart. Berman estimates the response speed at around 40–50 tokens per second in this session; that is an observation, not a measured service guarantee.

The generated chart gives Kimi K2.5 prices of $0.60 per million input tokens and $3 per million output tokens, and Claude Opus 4.5 prices of $5 and $25 respectively. These are figures shown in a January 2026 model-generated comparison; pricing and conditions change, so readers should verify current provider rates before deciding.

I then ask how much memory local use requires. The answer shown estimates about 632 GB for one INT4 configuration; requirements vary with the weights, quantization, runtime, context and offloading. K2.5 has about one trillion parameters, not one trillion tokens, with about 32 billion activated per token. Downloadable weights do not make the full model easy to run on ordinary hardware.

I hope the community will produce more practical quantized builds. The code and weights are released under a [modified MIT license](https://github.com/MoonshotAI/Kimi-K2.5/blob/master/LICENSE), rather than without conditions: commercial products or services above the specified monthly-user or revenue thresholds must prominently display the Kimi K2.5 name. I am keen to test the model and see what people build.