Kimi K2.7 Code review: coding and vision tests

Kimi K2.7 Code gets a mixed assessment in this review of coding and vision tasks. The speaker describes a mixture-of-experts model with 1 trillion parameters, 32 billion active parameters, a 256K context window and vision support. He reports that its weights are available on Hugging Face and cites approximately 30% lower thinking-token usage than Kimi K2.6. These are release claims discussed in the video, rather than independently verified measurements.

The tests use a Kimi coding plan with OpenCode and separate prompts through OpenRouter. This is a coding assistant review, not a guide to running a local LLM: the video does not establish hardware requirements or demonstrate local inference.

The speaker particularly likes the generated watch website and its 3D assets. A subway FPS includes environmental collision and bullet impacts, although enemy behavior remains questionable. The printer simulation accepts an uploaded STL and begins simulating its layers, despite poor rendering and slow progress.

Other results expose limits. A laptop reconstructed from photos has inverted geometry and Z-fighting. The demolition derby requires repeated fixes and still earns a failing assessment. Browser QA attempts repeatedly stall, which the speaker says could relate to his setup. He finds improvements over his earlier K2.6 tests, especially in the subway and skateboarding tasks, while acknowledging that his recent Claude Fable tests raised his expectations.