Kimi K2.5: Features, Pricing, and Use Cases
Kimi K2.5 is an open-weight, native multimodal AI model from Moonshot AI. It combines text, image, and video understanding with thinking and non-thinking modes, tool use, and agent workflows. Its model card documents a 1 trillion-parameter mixture-of-experts design with 32 billion activated parameters and a 256K-token context window.
What Is Kimi K2.5
K2.5 builds on Kimi K2 Base through continued pretraining on mixed visual and text data. It accepts visual inputs and produces text, which makes it relevant to screenshot-to-code, interface analysis, visual debugging, and document-heavy agent tasks. The published weights use the Modified MIT License, so teams should read that exact license and third-party notices before deployment.
Agent Swarm at Launch
K2.5 introduced Moonshot's Agent Swarm experience. At launch, Moonshot described a system that could create up to 100 subagents and coordinate up to 1,500 tool calls for work that can be split into parallel branches. That was a product and orchestration capability, not proof that every local K2.5 deployment automatically provides the same swarm.
The current Kimi Agent Swarm has moved to newer Kimi models. Treat K2.5's swarm figures as launch history, and evaluate the current product separately.
Kimi K2.5 vs DeepSeek V3.2 vs Claude
These models differ in more than benchmark scores. Use the exact model ID, reasoning mode, tool harness, and deployment route in any comparison.
| Decision factor | Kimi K2.5 | DeepSeek V3.2 | Claude |
|---|---|---|---|
| Weights | Published with Modified MIT terms | Check the named release license | Hosted model service |
| Modalities | Text, image, and video input | Verify the named endpoint | Text and image input on current models |
| Context | 256K | Verify the named release | Model-specific |
| Strong evaluation task | Visual coding and agent workflows | General or reasoning-mode coding | Long-horizon coding and tool work |
Coding Benchmarks
Moonshot publishes K2.5 results for coding, reasoning, vision, and agentic tasks, with test settings in the model card. Those results are vendor evidence, not a guarantee for a private repository. A useful pilot includes:
- Convert one approved UI image into a component and compare it with the design spec.
- Diagnose a failing integration test with access to only the relevant files and logs.
- Refactor one module, run the test suite, and report every changed public interface.
Record tool-call success, test pass rate, unnecessary edits, review time, and GPU or API cost.
Current Ecosystem and Access
K2.5 was released through Kimi's app, API, Kimi Code, and open weights. The current Kimi API catalog states that K2.5 is no longer available to newly registered users and is scheduled for a full platform sunset on August 31, 2026. It also lists Kimi K3 and K2.7 Code as newer options.
No first-party source currently establishes the model as the backbone of Cursor Composer 2; treat them as separate products. If platform access is unavailable, the published weights remain a separate deployment path, subject to the license and substantial infrastructure requirements.
Using Kimi K2.5 in Verdent
K2.5 is not listed in Verdent's current built-in model catalog. Verdent does document a BYOA configuration that can route Claude Code-compatible Kimi or Moonshot providers to this model. This uses your external provider endpoint and credentials; it is not the same as native Verdent model availability. Confirm that your provider account still permits access before configuring it.
Pilot One Agent Task Before Committing
Use a bounded visual or repository task, keep the acceptance criteria fixed, and measure review effort alongside model output.
Review Kimi provider setup in Verdent Verify Kimi availability in Verdent's catalog
Frequently Asked Questions
Is Kimi K2.5 an open-weight model?
Yes. Moonshot publishes Kimi K2.5 weights and code under a Modified MIT License. Review the license and third-party notices before commercial use or redistribution.
What is the Kimi K2.5 context window?
The official model card documents a 256K-token context window. Input modality, output allocation, memory, and serving configuration still affect practical use.
Does every Kimi K2.5 deployment include Agent Swarm?
No. Agent Swarm is an orchestration experience around the model. A local weight deployment needs its own agent harness, tools, permissions, and coordination layer.
Can I use Kimi K2.5 with Verdent?
Verdent documents Kimi K2.5 through an external BYOA provider route, not its current built-in catalog. Provider access is being sunset, so verify your account and endpoint first.