Post by AIME

Moonshot AI has released Kimi K2.7 Code today. The new model in the K2 family is specifically optimized for agentic code generation and long-horizon software engineering workflows. Technical Specifications: • Architecture: Mixture-of-Experts (MoE) with 1 Trillion (1T) parameters, of which 32 Billion are active per token. • Context Window: 256K tokens with automatic context compression. • License: Modified MIT (free for commercial use). • Availability: Open weights on Hugging Face, as well as via the Moonshot API. Key Improvements Compared to the Predecessor K2.6: - Efficiency Gains (Anti-Overthinking): The model requires approximately 30% fewer reasoning tokens for the same tasks. This significantly reduces inference costs and latency in production agentic workflows. - Multi-Language Performance: Beyond Python, the model shows significant progress in other languages. On the MLS Bench Lite (Python, Rust, Go), performance has improved by 31.5%. - Agentic Workflows: The preserve_thinking mode is enabled by default, retaining the full reasoning chain across multiple interactions, which increases reliability for complex, multi-step tasks. - Multimodality & Hardware Optimization: The model supports image and video inputs alongside text, and offers native INT4 quantization. Conclusion: With K2.7 Code, Moonshot AI is introducing a dedicated coding SKU within the K2 lineup for the first time. The shift from pure capability scaling towards token efficiency and multi-language support is a strong signal for the productive use of LLMs in software development. You can already use Kimi K2.7-code affordably at AKI.IO on EU-hosted hardware without hyperscalers, in a GDPR-compliant manner, and with zero data retention! https://lnkd.in/dn6gq2PP