Kimi K2.7-Code: open-source coding model with better token efficiency (huggingface.co)

Kimi Technologies has launched Kimi K2.7 Code, an open-source coding model that greatly enhances token efficiency and performance on complex coding tasks compared to its predecessor, Kimi K2.6. The model utilizes a Mixture-of-Experts architecture with 1 trillion total parameters, activating 32 billion for improved task execution. Notably, it achieves a 30% reduction in thinking-token usage, which is pivotal for tackling long-horizon coding assignments prevalent in software engineering workflows. Benchmarks indicate significant improvements across various coding tasks, outperforming existing models such as GPT-5.5 and Claude Opus 4.8. This upgrade holds great significance for the AI/ML community as it enables more efficient coding automation and responsiveness in software development. The model is equipped with advanced features, including a 256k-token context length and the ability to handle multi-modal inputs, such as images and videos. Its native int4 quantization and strict adherence to high-performance benchmarks make it an attractive choice for developers looking to enhance productivity and reduce resource consumption in coding applications. Kimi K2.7 Code can be accessed via its API, reflecting a commitment to open collaboration in advancing AI-driven coding solutions.