Try Kimi K2.7 Code
256K context • coding agents • always-on thinking • text + image input
Kimi K2.7 Code Assistant
repository-scale coding • multi-step tools • image-aware debugging
Fast Response
Get instant answers powered by our optimized infrastructure
Privacy First
Your conversations are secure and never used for training
Premium Features
Sign in to unlock API access and unlimited conversations
Kimi K2.7 Code in four numbers
A compact view of the architecture and operating profile published by Moonshot AI for Kimi K2.7 Code.
Context Window
256K
Room for repository context, logs, specifications, and long agent sessions
MoE Architecture
1T / 32B
One trillion total parameters with 32 billion activated per token
Thinking Tokens
≈30% Less
Average reduction reported against K2.6 on three coding evaluations
Kimrel Minimum
3 Credits
Minimum charge per successful request before token-based billing applies
What makes Kimi K2.7 Code different
Kimi K2.7 Code is a coding-focused agentic model built on K2.6. Its improvements target the parts of software work that usually break a coding assistant: long instructions, repeated tool calls, evolving context, and end-to-end completion.
Long-horizon instruction fidelity
Kimi K2.7 Code is tuned to follow detailed instructions more reliably as a session grows. That matters when acceptance criteria, repository rules, and earlier decisions must survive many rounds of inspection, editing, testing, and correction.
End-to-end software work
The model is aimed at complete engineering workflows rather than isolated code completion. Kimi K2.7 Code can reason from a task, inspect evidence, propose changes, use tools, and revise its approach when tests or runtime output disagree.
Less unproductive overthinking
Moonshot AI reports that Kimi K2.7 Code uses about 30% fewer thinking tokens on average than K2.6 across three coding benchmarks. It remains a reasoning model, but it is trained to reach useful action with less internal wandering.
Interleaved thinking and tools
Kimi K2.7 Code supports multi-step tool calls with reasoning between actions. A coding agent can read a file, interpret the result, run another tool, and continue without reducing the task to one oversized prompt.
Reasoning that persists across turns
The official model card says preserve-thinking behavior is always enabled. Kimi K2.7 Code can retain prior reasoning context during a multi-turn coding workflow, which helps it stay consistent when a later request depends on earlier technical decisions.
Visual context for debugging
A 400M-parameter MoonViT encoder gives Kimi K2.7 Code native visual input. On Kimrel, developers can combine text with screenshots, diagrams, or UI captures when a bug cannot be explained by source code alone.
How Kimi K2.7 Code improves on K2.6
Moonshot AI evaluated Kimi K2.7 Code and K2.6 with thinking enabled in Kimi Code CLI. The published results point to a clear coding-agent gain, although benchmark scores should remain one input alongside tests on your own repositories.
Kimi Code Bench v2 — 62.0
Kimi K2.7 Code rises from K2.6's 50.9 to 62.0 on Moonshot AI's in-house benchmark covering production incidents and software tasks across more than ten programming languages.
Program Bench — 53.6
Kimi K2.7 Code scores 53.6 versus 48.3 for K2.6. Program Bench asks agents to recreate a program from its executable behavior and documentation, then checks it with fuzz-generated tests.
MLS Bench Lite — 35.1
Kimi K2.7 Code reaches 35.1, up from 26.7. This evaluation covers open-ended machine-learning systems work, including training, robotics, optimization, computer vision, and research-oriented tasks.
Kimi Claw 24/7 — 46.9
Kimi K2.7 Code improves from 42.9 to 46.9 on Moonshot AI's persistent-agent benchmark, which spans multi-day professional scenarios and tests whether an agent can keep working toward a durable outcome.
MCP Atlas — 76.0
Kimi K2.7 Code posts 76.0 versus 69.4 for K2.6 under the published configuration. MCP Atlas focuses on realistic tool use, making it especially relevant to developers building function-calling agents.
MCP Mark Verified — 81.1
Kimi K2.7 Code moves from 72.8 to 81.1 across verified tasks using GitHub, filesystem, Postgres, Notion, and Playwright environments. The result supports its positioning as a practical tool-using coding model.
Where Kimi K2.7 Code earns its place
Choose Kimi K2.7 Code when the job is fundamentally software engineering and the model must stay useful beyond the first answer. These workflows make direct use of its long context, persistent reasoning, tools, and vision.
Repository-wide feature delivery
Give Kimi K2.7 Code the feature brief, relevant architecture, coding conventions, and test expectations. It is suited to work that crosses routes, services, types, migrations, and verification instead of ending with a single function.
Framework upgrades and migrations
Kimi K2.7 Code can keep migration constraints in view while tracing affected files, adjusting interfaces, and responding to build failures. Use staged checks so each change is grounded in the repository's actual behavior.
Evidence-led debugging
Combine stack traces, logs, source excerpts, and screenshots. Kimi K2.7 Code can connect symptoms across artifacts, form a testable hypothesis, and refine it after commands or diagnostic tools return new evidence.
Tool-driven coding agents
Use Kimi K2.7 Code in loops that read files, search symbols, run tests, call functions, and inspect results. Interleaved thinking is most valuable when each tool response changes the next decision.
Frontend work from screenshots
Send a UI capture with the implementation context and a precise objective. Kimi K2.7 Code can identify visible hierarchy, missing states, spacing problems, or responsive defects before proposing component-level changes.
Test and review assistance
Kimi K2.7 Code can map requirements to test cases, inspect risky diffs, and explain failure paths. Human review still matters, especially for security, data migrations, permissions, and production-facing changes.
Use Kimi K2.7 Code through Kimrel
Kimrel provides independent access to Kimi K2.7 Code through familiar API shapes. The service is not affiliated with Moonshot AI and documents its own billing and multimodal boundaries separately from the underlying model.
Select the exact model ID
Set `model` to `kimi-k2.7-code`. Kimi K2.7 Code uses the same Kimrel API key flow as other supported routes, so an existing integration only needs an explicit model selection.
Keep thinking enabled
Kimi K2.7 Code does not offer a non-thinking mode. Leave enough output budget for both reasoning and the final answer; an overly small completion limit can exhaust the response before useful content appears.
Use either compatible route
Call Kimi K2.7 Code through Kimrel's OpenAI-compatible `/v1/chat/completions` endpoint or Anthropic-compatible `/v1/messages` endpoint. Follow the API documentation for each route's authentication and payload shape.
Send text and images
Kimrel enables text and image input for Kimi K2.7 Code. The upstream model also documents video capability, but video is intentionally not accepted by this service and should not be included in requests.
Let Kimrel encode remote images
For Kimi K2.7 Code, an HTTP(S) image URL can be fetched, validated, and converted to base64 before forwarding. Kimrel limits the original file to 6MB and the encoded payload to 8MB.
Plan for token-based credits
Kimi K2.7 Code is configured at 95 input credits and 400 output credits per one million tokens, with a three-credit minimum per successful request. Actual wallet usage follows Kimrel's metering rules.
Kimi K2.7 Code questions, answered
Short answers to the practical questions developers ask before choosing this coding model.
What is Kimi K2.7 Code?
Kimi K2.7 Code is a coding-focused agentic model developed by Moonshot AI on top of K2.6. It targets long-horizon software engineering, stronger instruction following, multi-step tool use, and higher end-to-end completion rates. Kimrel provides independent access and is not affiliated with Moonshot AI.
Is it a general-purpose replacement for K2.6?
Not necessarily. Kimi K2.7 Code is the better fit when code and agent execution dominate the task. Moonshot AI recommends K2.6 for broader writing, analysis, and conversation, so model choice should follow the workload rather than the version number.
Does Kimi K2.7 Code always use thinking?
Yes. Official documentation says Kimi K2.7 Code forces thinking and preserve-thinking behavior. Set a realistic completion budget, especially for complex repository requests, because reasoning tokens are part of the generated output budget before the final response is complete.
Can Kimi K2.7 Code understand screenshots?
Yes. Kimi K2.7 Code includes the MoonViT vision encoder. Kimrel accepts base64 image data or remote HTTP(S) image URLs for this route, then forwards a validated base64 payload upstream. Video input remains unavailable on Kimrel.
Does it support function and tool calls?
Yes. Kimi K2.7 Code supports interleaved thinking and multi-step tool calling. It is designed for agents that inspect one result before choosing the next action, but callers should still validate arguments, limit permissions, and handle tool failures explicitly.
Are the published benchmark scores guaranteed in my app?
No. The Kimi K2.7 Code scores describe Moonshot AI's published test configurations, not a guarantee for every repository or agent harness. Evaluate it with representative tasks, the same tools used in production, and checks that measure correctness rather than output style.
Put Kimi K2.7 Code on a real coding task
Try Kimi K2.7 Code in chat or use the Kimrel API for repository-scale, tool-assisted engineering workflows.