Kimi K2.7 Code on Run BiOS

Kimi K2.7 Code is an open-source, coding-focused agentic model developed by Moonshot AI, built upon Kimi K2.6. Moonshot AI reports substantial improvements on real-world long-horizon coding tasks: it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.

Highlights reported by Moonshot AI:

• Long-horizon coding. Higher success rates on sequences of 4,000+ tool calls and 12+ hours of continuous execution, with generalization across languages (Rust, Go, Python) and tasks (frontend, DevOps, performance optimization). Compared with K2.6, the model follows instructions more reliably in long contexts and achieves higher end-to-end task success rates.
• Optimized reasoning efficiency. Reasoning models tend to overthink, spending thousands of tokens deliberating on problems that do not need it. Moonshot AI reports that Kimi K2.7 Code significantly reduces this tendency, cutting thinking-token usage by approximately 30% on average compared with K2.6 — and that across Kimi Code Bench v2, Program Bench, and MLS Bench Lite it achieves higher scores than K2.6 while consuming fewer tokens on each benchmark.
• Multimodal tool use. Native video ingestion with temporal analysis and clip extraction via custom tools. Images (PNG, JPEG, WebP, GIF) and video (MP4, MOV, AVI, and other common formats) are supported, per Moonshot AI's release announcement; the model card notes that chat with video content is an experimental feature.
• Reproducible outputs. Temperature (1.0), top-p (0.95), n (1), and presence/frequency penalties (0.0) are fixed at the vendor-recommended values for consistent, reproducible outputs, per Moonshot AI's release announcement.
• Positioning. Moonshot AI describes Kimi K2.7 Code as purpose-built for coding tasks; for general-purpose work such as writing, analysis, and conversation, the company recommends K2.6, which it describes as more well-rounded.

Architecture

All architecture figures are from Moonshot AI's official model card.

Property	Value
Architecture	Mixture-of-Experts (MoE)
Total Parameters	1T
Activated Parameters	32B
Number of Layers (Dense layer included)	61
Number of Dense Layers	1
Attention Hidden Dimension	7168
MoE Hidden Dimension (per Expert)	2048
Number of Attention Heads	64
Number of Experts	384
Selected Experts per Token	8
Number of Shared Experts	1
Vocabulary Size	160K
Context Length	256K
Attention Mechanism	MLA
Activation Function	SwiGLU
Vision Encoder	MoonViT
Parameters of Vision Encoder	400M

Kimi K2.7 Code shares the same architecture as Kimi K2.5 and Kimi K2.6, per Moonshot AI. It uses Multi-head Latent Attention (MLA) and includes MoonViT, a 400M-parameter vision encoder. The model adopts the same native INT4 quantization method as Kimi K2 Thinking. Moonshot AI's published API configuration for the model specifies a 262,144-token context window (256K).

Benchmarks

All results below are from Moonshot AI's official model card. Kimi K2.7 Code was evaluated against K2.6 on a combination of internal and external benchmarks covering two dimensions: coding capability and agentic task execution. Moonshot AI reports coding gains over K2.6 of +21.8% on Kimi Code Bench v2 (62.0 vs 50.9), +11.0% on Program Bench (53.6 vs 48.3), and +31.5% on MLS Bench Lite (35.1 vs 26.7), and agentic gains of roughly 10% over K2.6 on Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified.

Coding

Benchmark	Kimi K2.6	Kimi K2.7 Code	GPT-5.5	Claude Opus 4.8
Kimi Code Bench v2	50.9	62.0	69.0	67.4
Program Bench	48.3	53.6	69.1	63.8
MLS Bench Lite	26.7	35.1	35.5	42.8

Agentic

Benchmark	Kimi K2.6	Kimi K2.7 Code	GPT-5.5	Claude Opus 4.8
Kimi Claw 24/7 Bench	42.9	46.9	52.8	50.4
MCP Atlas	69.4	76.0	79.4	81.3
MCP Mark Verified	72.8	81.1	92.9	76.4

Benchmark methodology notes (as disclosed by Moonshot AI)

• General testing details. Unless stated otherwise, Kimi K2.7 Code and K2.6 were tested with thinking mode enabled via the Kimi Code CLI at temperature = 1.0, top-p = 0.95, and a 262,144-token context length; GPT-5.5 ran in Codex with xhigh mode, and Opus 4.8 in Claude Code with xhigh mode. Aside from these differences, all benchmarks were evaluated under the same conditions.
• Kimi Code Bench v2 (in-house). Moonshot AI's in-house benchmark designed to evaluate coding agents on realistic tasks. It has diverse software-engineering tasks across 10+ mainstream programming languages and a full production tech stack covering tasks from internal engineering use cases, production incidents, and real-world open-source projects, with emphasis on backend services, infrastructure, performance engineering, systems programming, security, frontend development, and ML/data engineering.
• Program Bench. Evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and its documentation. It spans 200 tasks, from small CLI tools to large systems like FFmpeg and SQLite. Submissions are judged against over 248,000 fuzz-generated behavioral tests. In each task, the agent is given an executable and its documentation, but no source code, decompilation, or internet access. It must choose its own implementation language, build the full program from scratch, and pass a behavioral test suite comparing its output against the original binary.
• MLS Bench Lite. MLS-Bench evaluates whether AI systems can invent generalizable and scalable ML methods. MLS-Bench-Lite is the official 30-task subset of MLS-Bench, covering LLM pretraining and post-training, robotics, world models, computer vision, reinforcement learning, optimization, ML systems, AI for Science, and more. Agents are given 5 hours to explore before submitting their solutions. Opus 4.8 is evaluated with the max effort setting in Claude Code.
• Kimi Claw 24/7 Bench (in-house). Moonshot AI's in-house benchmark for evaluating long-horizon agentic performance in persistent, multi-day coworking tasks. It spans 17 professional scenarios across 610 evaluation points, covering domains such as software engineering, ML research, recruiting, trading, and marketing. All tasks are executed through the OpenClaw harness. The final score is the average pass rate across all evaluation points, averaged over 3 runs.
• MCP Atlas. Evaluates LLM performance on realistic tool-use tasks through scalable MCPs. Moonshot AI followed the official MCP-Atlas evaluation configuration with a 100 tool-call budget and 32k max tokens per step. The final result is averaged over 3 runs.
• MCP Mark Verified. A human-verified edition of MCPMark, a benchmark for evaluating MCP tool use across five real server environments — Notion, a code-hosting service, a filesystem, Postgres, and a browser-automation environment. Each task has been re-checked by Moonshot AI's team and the benchmark's authors. Moonshot AI followed the official MCPMark evaluation configuration with a 100-step tool-call budget and 32k max tokens per step. The final result is averaged over 3 runs.

Thinking and Tool Use on Run BiOS

• Reasoning is always on. Per Moonshot AI's documentation, Kimi K2.7 Code does not support a non-thinking mode: it always runs with thinking enabled, and there is no reasoning-effort ladder to configure. It also forces preserve-thinking mode, which retains full reasoning content across multi-turn interactions and enhances performance in coding-agent scenarios; this behavior is enabled by default and cannot be disabled.
• Function calling. Kimi K2.7 Code supports interleaved thinking and multi-step tool calls, sharing the same design as Kimi K2 Thinking, per Moonshot AI's model card.
• Multimodal input. The model supports image and video input in addition to text (video chat is described by Moonshot AI as an experimental feature).
• Context window. 262,144 tokens (256K), matching the published evaluation context length.
• Prompt caching. Repeated prefixes are served from cache at a reduced rate.

Safety and Intended Use

• Intended use. Moonshot AI positions Kimi K2.7 Code as purpose-built for coding and agentic software-engineering workflows; for general-purpose writing, analysis, and conversation, the company recommends Kimi K2.6 instead.
• Deterministic behavior. Fixed sampling parameters (temperature 1.0, top-p 0.95, n 1, zero presence/frequency penalties) are locked by the vendor for consistent, reproducible outputs.
• Thinking-mode requirement. Because thinking is always enabled, requests that attempt to disable thinking are not served by this model in Moonshot AI's own stack (they are routed to K2.6 there). On Run BiOS the model simply always responds with thinking enabled.