GoKawiil1. Model Introduction
Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With 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.
2. Model Summary
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
3. Evaluation Results
Benchmark Kimi K2.6 Kimi K2.7 Code GPT-5.5 Claude Opus 4.8 Coding 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 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
Footnotes General Testing Details Unless stated otherwise, Kimi K2.7 Code and K2.6 were tested with thinking mode enabled via 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. Coding Benchmarks Kimi Code Bench V2 is our in-house benchmark designed to evaluate coding agents on realistic tasks. It has diversed 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 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. Agentic Benchmarks Kimi Claw 24/7 Bench is our 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, marketing. All tasks are executed through the OpenClaw harness. The final score is the average pass rate across all evaluation points, and is averaged over 3 runs.
MCP-Atlas evaluates LLM performance on realistic tool-use tasks through the scalable MCPs. We followed the official MCP-Atlas evaluation configuration with a 100 tool-call budget, and with 32k max tokens per step. The final result is averaged over 3 runs.
... continue reading