Agentic software development
Long, multi-step tasks that require reading a repository, editing files, running tools and iterating on failures.
Independent tool overview
Kimi K2.7 Code at a glance
Kimi K2.7 Code is Moonshot AI's open-weight, multimodal coding model for long-horizon software tasks, tool use and agent workflows, available through Kimi Code, a pay-as-you-go API and self-hosted deployment.
Visit the official Kimi K2.7 Code site ↗
- Developer
- Moonshot AI
- Architecture
- 1T-parameter mixture of experts, 32B active
- Context window
- 256K tokens
- Inputs
- Text, images and video in supported access paths
- Reasoning
- Thinking-only with preserved reasoning across turns
- License
- Modified MIT for code and model weights
Overview
What Kimi K2.7 Code is
Kimi K2.7 Code is a coding-focused mixture-of-experts model built on Kimi K2.6. It has one trillion total parameters, activates 32 billion per token, supports a 256,000-token context and accepts text, image and video input in supported environments.
Moonshot positions the model for end-to-end software engineering and reports about 30% lower thinking-token usage than K2.6. Its model card shows strong gains on Moonshot's coding and tool-use evaluations, while several comparison benchmarks are internal or use different agent harnesses and reasoning settings. Treat those numbers as vendor evidence to reproduce on your own repositories.
There are three distinct ways to use it: Kimi Code subscriptions for terminal, VS Code and compatible coding clients; the Kimi API with token-based billing; or downloaded weights under a Modified MIT License. Self-hosting is technically possible but the trillion-parameter model is infrastructure-heavy—the official guide's standard examples use eight H200 GPUs.
Use cases
Who Kimi K2.7 Code is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Cost-conscious API teams
Developers comparing capable coding models at published per-token prices.
Large self-hosting programs
Organizations with substantial GPU or heterogeneous inference infrastructure and a reason to control model deployment.
Capabilities
Core Kimi K2.7 Code features
Long-context coding
A 256K-token context supports large code excerpts, tool histories and multi-step agent sessions.
Agentic tool use
The model supports interleaved reasoning and multi-step tool calls for coding agents and MCP-style environments.
Multimodal understanding
Image and video inputs can supply screenshots, diagrams and visual application context; video support is limited by the access path.
Preserved reasoning
Prior reasoning content is retained across turns to improve continuity in long coding workflows, and the model card says this mode cannot be disabled.
Native INT4 weights
Moonshot publishes native INT4 quantization and deployment guidance for vLLM, SGLang and KTransformers.
Kimi Code clients
Use the model in the official terminal and VS Code tools or connect supported third-party coding clients with a membership key.
OpenAI and Anthropic compatibility
Official API and coding endpoints provide familiar interfaces for integrating the model into existing developer tooling.
Process
How the Kimi K2.7 Code workflow works
Step 1
Choose an access path
Use Kimi Code for an integrated coding agent, the token API for an application, or self-hosting only when control justifies the infrastructure.
Step 2
Start with a bounded repository task
Give the agent clear acceptance criteria, relevant tests and the smallest permissions needed to inspect and modify the code.
Step 3
Preserve useful context
Let the model inspect repository instructions and tool output, but compact or restart sessions when stale context begins to outweigh useful evidence.
Step 4
Review every material change
Inspect diffs, run automated tests, scan dependencies and verify security-sensitive logic before merging or deploying.
Step 5
Measure model economics
Track cache hit rate, input and reasoning growth, output tokens, latency and successful task completion rather than comparing headline token prices alone.
Cost
Kimi K2.7 Code pricing and free plan
Downloaded weights have no usage fee but require expensive infrastructure and license compliance. The official Kimi API bills cached input, uncached input and output separately. Kimi Code is also bundled into RMB-priced Kimi memberships with shared credits and rate limits.
Open weights
Free to download
Self-host Kimi K2.7 Code under Moonshot's Modified MIT License.
- Compute, storage and operations are separate
- Official full-model examples use 8 H200 GPUs
- Large commercial services may have a UI attribution requirement
Kimi API
$0.19 cache / $0.95 input / $4 output per 1M tokens
Pay-as-you-go official model inference.
- 256K context
- OpenAI-compatible access
- Both input and output are billed
Andante
¥49 per month
Entry Kimi membership with standard Kimi Code access.
- Shared membership credits
- Kimi Code Standard available
- Unused monthly credits expire
Moderato
¥99 per month
Higher shared credits and productivity features.
- Kimi Code Standard available
- Additional parallel agent and Kimi product capacity
Allegretto
¥199 per month
Professional membership that unlocks K2.7 Code HighSpeed.
- HighSpeed is advertised at 5–6x output speed
- HighSpeed uses about 3x the quota
- Shared credits and rate limits apply
Allegro
¥699 per month
The highest published individual Kimi membership.
- Largest shared credit pool
- Kimi Code Standard and HighSpeed access
- Other Kimi product benefits share the pool
Third-party inference
Varies by provider
Hosted access is also listed through Hugging Face inference partners.
- Provider pricing, limits and feature support differ
Pricing checked . Check current pricing at the source ↗
Assessment
Kimi K2.7 Code strengths and limitations
Where it stands out
- Publishes weights and detailed deployment guidance in addition to managed access.
- The official API has straightforward and competitive token prices for a coding-focused model.
- A 256K context and preserved reasoning suit long agent sessions and tool-heavy tasks.
- Native INT4 and multiple inference-engine options make a very large model somewhat more deployable.
- Kimi Code supports official clients and compatible third-party coding tools.
What to consider
- At one trillion total parameters, self-hosting remains a data-center-scale project despite 32 billion active parameters and INT4 weights.
- Thinking and preserve-thinking modes cannot be disabled, which increases token and context use and requires careful handling of reasoning data.
- The reported 30% thinking-token reduction is relative to Kimi K2.6 and does not guarantee a lower total bill than another model.
- Several highlighted evaluations are created or re-verified by Moonshot, and competing models used different tools or effort settings.
- Video input is experimental outside the official API and may not work through every self-hosted or third-party endpoint.
- Kimi memberships share credits across products and impose additional rate or quota windows; a listed monthly price is not unlimited coding use.
- The Modified MIT License requires prominent Kimi K2.7 Code identification when a commercial product exceeds 100 million monthly users or $20 million in monthly revenue.
- Like any coding agent, it can introduce defects, unsafe commands or vulnerable code and must be constrained, reviewed and tested.
Compare
Kimi K2.7 Code alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Coding
Qwen3-Coder
Choose Qwen3-Coder when comparing another open-weight mixture-of-experts family built specifically for agentic coding.
Explore Qwen3-Coder →Coding
GLM-4.7
Choose GLM-4.7 for a different open-weight reasoning and coding stack with its own hosting and API options.
Explore GLM-4.7 →Coding
Claude Code
Choose Claude Code when a mature managed coding-agent product matters more than downloadable model weights.
Explore Claude Code →Questions
Kimi K2.7 Code FAQs
Is Kimi K2.7 Code open source?
Moonshot publishes the model weights and related code under a Modified MIT License. The modification adds a branding requirement for very large commercial products, so review the actual license before distribution.
How much does the Kimi K2.7 Code API cost?
The current official platform lists $0.19 per million cache-hit tokens, $0.95 per million uncached input tokens and $4 per million output tokens.
Can I run Kimi K2.7 Code locally?
The weights can be self-hosted, but this is not a typical laptop model. Moonshot's straightforward vLLM and SGLang examples use tensor parallelism across eight H200 GPUs, with a more complex CPU-GPU route also documented.
What is the difference between Kimi Code and Kimi K2.7 Code?
Kimi K2.7 Code is the model. Kimi Code is Moonshot's coding-agent service and clients, which can use K2.7 Code or newer supported Kimi models through a membership subscription.
Does Kimi K2.7 Code support images and video?
Yes, the model card lists both. Image input works in official examples, while video is currently an official-API feature and experimental or unavailable in some third-party deployments.
Is the model 30% cheaper than Kimi K2.6?
Moonshot says it uses about 30% fewer thinking tokens than K2.6. That is not the same as 30% lower total cost because input, output, cache behavior, task success and access pricing also matter.
Bottom line
Our Kimi K2.7 Code verdict
Kimi K2.7 Code is a credible option for teams seeking a strong coding model with open weights and low published API rates. The managed API or Kimi Code service will be the practical route for most users; self-hosting only makes sense at serious scale. Evaluate it on private tasks, account for forced reasoning and verify every code change rather than buying from vendor benchmarks alone.
Visit Kimi K2.7 Code website ↗Similar Coding Tools
View all Coding AI tools →
North Mini Code
⚙️ North Mini Code - Cohere's first open-source agentic coding model
GLM 5.2
GLM 5.2 - Z AI's new flagship coding model with usable 1M context
Devin Desktop
Devin Desktop - Cognition's Windsurf rebrand for managing all your coding agents in a single platform
eve
eve - Vercel's open-source framework that turns a file directory into an agent
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.