Multimodal coding
Turn screenshots, visual references, and video into front-end code and use visual feedback during debugging.
Independent tool overview
Kimi K2.5 at a glance
Kimi K2.5 is Moonshot AI's 1-trillion-parameter multimodal agentic model with a 256K context window, open weights, vision and video input, coding capabilities, and optional multi-agent orchestration.
Visit the official Kimi K2.5 site ↗
- Developer
- Moonshot AI
- Architecture
- 1T MoE; 32B active parameters
- Context
- 262,144 tokens
- Inputs
- Text, images, and official-API video
- License
- Modified MIT
Overview
What Kimi K2.5 is
Moonshot AI released Kimi K2.5 in January 2026 as a native multimodal model for reasoning, coding, visual understanding, tool use, and knowledge work. The mixture-of-experts model has 1 trillion total parameters, activates 32 billion per token, and supports a 256K context window.
K2.5 can operate in instant or thinking modes and accept text, image, and—through the official API—experimental video input. Its Agent mode can create documents, spreadsheets, PDFs, slides, and software, while Agent Swarm can decompose a large task across dynamically created subagents. Moonshot's published swarm scale and benchmark gains are vendor-reported and should be reproduced on representative work before adoption.
K2.5 remains listed in the Kimi API, but it is no longer the flagship. Kimi K2.6 followed with stronger agentic coding and long-horizon work, and Kimi K3 is now Moonshot's most capable model with a 1-million-token context window. New integrations should compare all three instead of assuming the older model is the best current default.
Use cases
Who Kimi K2.5 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Long-context analysis
Work across large document, image, code, or mixed-media inputs within a 256K-token context window.
Office deliverables
Generate structured documents, spreadsheets, slide decks, PDFs, and other end-to-end knowledge-work outputs.
Parallel research
Use Agent Swarm for large search, reading, downloading, coding, and batch-processing tasks that can be decomposed safely.
Capabilities
Core Kimi K2.5 features
Native multimodality
The model was trained jointly on visual and text data and can reason over text, images, and supported video inputs.
Instant and thinking modes
Choose lower-latency interaction or deeper reasoning depending on the task and cost profile.
Coding with vision
Generate and refine interfaces from screenshots or video and combine coding with visual inspection.
Agent tool use
Use tools across multi-step research, development, data, and office workflows.
Agent Swarm
Moonshot's beta orchestration mode can dynamically create specialized subagents and run parallel work.
Open weights
Model weights and code are available under a Modified MIT license with an attribution condition for very large commercial services.
Process
How the Kimi K2.5 workflow works
Step 1
Choose the right Kimi generation
Benchmark K2.5 against K2.6 and K3 on quality, latency, context, modalities, and price before locking the model ID.
Step 2
Select the mode
Use instant mode for straightforward tasks, thinking mode for difficult reasoning, and agents only when tools materially improve the result.
Step 3
Constrain tool access
Give agents narrow permissions, isolated environments, spending caps, and explicit approval points for external actions.
Step 4
Validate outputs
Run tests for code, check citations and calculations, inspect generated files, and compare vendor benchmarks with internal evaluations.
Step 5
Track cache and tokens
Monitor cached versus uncached input because official API prices differ substantially and long contexts can dominate cost.
Cost
Kimi K2.5 pricing and free plan
Kimi API pricing is denominated in Chinese yuan per million tokens. K2.5 remains cheaper than the newer K3 flagship, particularly for uncached input and output.
Kimi K2.5 cached input
¥0.70 / 1M tokens
Input price when automatic context caching hits.
- 256K context
- Usage based
- Cache behavior affects cost
Kimi K2.5 uncached input
¥4.00 / 1M tokens
Standard input price when the prompt is not served from cache.
- Text, image, and supported video input
- Usage based
- Official API
Kimi K2.5 output
¥21.00 / 1M tokens
Generated-token price for K2.5 API responses.
- Thinking can increase output volume
- Tool loops add usage
- Usage based
Kimi K3
¥2 cached / ¥20 input / ¥100 output per 1M
Current flagship pricing for teams comparing the official successor.
- 1M-token context
- Native vision
- Higher cost than K2.5
Pricing checked . Check current pricing at the source ↗
Assessment
Kimi K2.5 strengths and limitations
Where it stands out
- Strong combination of vision, coding, tool use, and long-context reasoning
- Open weights create self-hosting and research options beyond the official API
- Official API pricing is low relative to many frontier hosted models
- Agent and office workflows extend beyond chat into finished files and applications
- K2.5 remains available even after K2.6 and K3 expanded the model family
What to consider
- K2.5 is no longer Moonshot's most capable model, so new integrations should evaluate K2.6 and K3
- Self-hosting a 1T-parameter MoE model requires substantial specialized infrastructure despite only 32B active parameters
- Agent Swarm is a complex beta capability whose cost, reliability, and permission risks grow with parallelism
- Moonshot's benchmark and time-saving claims are vendor-reported and may not transfer to real workloads
- The Modified MIT license adds a display requirement for commercial services above specified scale thresholds
- Experimental video support and some behaviors may differ between the official API and third-party hosts
Compare
Kimi K2.5 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
DeepSeek
A competing open-model ecosystem with strong reasoning and coding options.
Explore DeepSeek →Agents
Mistral AI
A broader open-weight and hosted model platform with enterprise deployment choices.
Explore Mistral AI →Business Operations
ChatGPT
A more polished general-purpose product for users who prefer a managed assistant over model deployment.
Explore ChatGPT →Questions
Kimi K2.5 FAQs
Is Kimi K2.5 still available?
Yes. Moonshot's current API documentation still lists the `kimi-k2.5` model, although K2.6 and K3 are newer options.
What replaced Kimi K2.5?
Kimi K2.6 was the direct upgrade for agentic coding and long-horizon execution. Kimi K3 is now Moonshot's most capable flagship model.
Is Kimi K2.5 open source?
Moonshot publishes the model weights and code under a Modified MIT license. Large commercial services above the license thresholds must prominently display the Kimi K2.5 name.
How much does the K2.5 API cost?
Moonshot lists ¥0.70 per million cached input tokens, ¥4.00 per million uncached input tokens, and ¥21.00 per million output tokens.
What is Kimi Agent Swarm?
It is a parallel orchestration mode that can dynamically create specialized subagents for decomposable research, coding, and office tasks. It remains a complex capability that needs tight permissions and cost controls.
Can K2.5 understand video?
Yes, but Moonshot labels chat with video as experimental and says it is supported through the official API; third-party deployments may differ.
Bottom line
Our Kimi K2.5 verdict
Kimi K2.5 remains a capable and inexpensive multimodal agent model with open weights, especially for vision-assisted coding and structured knowledge work. It should now be evaluated as a cost-conscious member of the Kimi family rather than the default flagship, with K2.6 and K3 included in every new-model bakeoff.
Visit Kimi K2.5 website ↗Similar Agents Tools
View all Agents AI tools →
MiniMax Agent
MiniMax Agent - AI assistant with full browser control, expert agent skills, and computer and cloud capabilities
OpenAI Frontier
OpenAI Frontier - enterprise platform for creating, deploying, and managing AI agents
OpenClaw
OpenClaw - Agentic AI bot that takes actions for you from Telegram, WhatsApp, or other messaging apps
Kimi Claw
Kimi Claw- Moonshot's browser-based AI assistant that deploys integrated OpenClaw agents in the cloud with 5,000+ community skills
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.