The latest AI news and updates of Kimi — an AI assistant and open model family spanning chat, coding, agents, and vision. Covering Kimi's latest launches, product updates, and research from the past 90 days.
Kimi14h agoMoonshot AI is offering a promotional credit bonus for Kimi API developers building with the new Kimi K2.7 Code model. Through July 2, 2026, single top-ups of $100 or more earn a tiered voucher bonus ranging from 20% to 30%. These vouchers are valid for 90 days and capped at $4,000 per organization.
Kimi15h agoMoonshot AI released Kimi-K2.7-Code, an open-source coding model featuring significant benchmark gains, including a 21.8% improvement on Kimi Code Bench v2. The model achieves higher performance while using 30% fewer reasoning tokens than K2.6. It is available today via the Kimi API and Kimi Code, with a new beta program and high-speed mode launching soon.
KimiJun 9Moonshot AI has launched Kimi Work, a local desktop AI agent for macOS and Windows. It features a native agent swarm of up to 300 parallel agents, browser automation, and integrated finance tools to autonomously perform tasks and learn user preferences. This brings advanced agentic capabilities directly to the user's machine for continuous, personalized workflow automation.
KimiMay 15Moonshot AI released Kimi Web Bridge, a browser extension that allows AI agents to interact with websites by clicking, typing, and scrolling. The tool is interoperable, enabling terminal-based agents like Claude Code and Cursor to use the browser for research and task execution within the user's authenticated session.
KimiApr 27Kimi K2.6 reached the #1 position on OpenRouter's weekly LLM leaderboard, specifically outperforming established frontier models in programming tasks. This third-party validation confirms the model's real-world utility for developers just days after its release.
KimiApr 24Kimi K2.6 is now the highest-ranked model for programming on OpenRouter, outperforming established frontier models in coding tasks. This third-party validation confirms the model's effectiveness for complex software development workflows just days after its release.
KimiApr 24Moonshot AI released Kimi K2.6 Agent Swarm, a system that coordinates up to 300 parallel sub-agents across 4,000 steps to complete complex tasks. The update shifts from conversational responses to production-ready file bundles, including 100-page reports and massive datasets.
KimiApr 21Kimi open-sourced FlashKDA, a high-performance implementation of its Kimi Delta Attention kernels built on the CUTLASS library. The release achieves up to a 2.22x speedup in prefill processing on H20 GPUs compared to standard linear attention baselines. This allows developers to significantly reduce latency for long-context applications by swapping in a more efficient backend.
KimiApr 20Moonshot AI released Kimi K2.6, an open-source model that achieves state-of-the-art scores on SWE-Bench Pro and Toolathlon. The update introduces long-horizon coding, enabling agents to execute over 4,000 autonomous steps without losing context or drifting from the task.
KimiApr 18Moonshot AI launched Prefill-as-a-Service, a distributed architecture that separates the compute-heavy prefill phase from the memory-heavy decode phase across different datacenters. By using a hybrid model to shrink memory overhead, the system achieves 1.54x higher throughput and significantly lower latency for long-context requests.
KimiMar 21Kimi CEO Zhilin Yang detailed the training innovations behind Kimi K2.5 at NVIDIA GTC 2026. The session covers the Muon optimizer replacing Adam to double token learning efficiency, AI-native training, and a shift toward linear attention for longer-running agents.
KimiMar 21Moonshot AI confirmed Kimi-k2.5 is the foundation model behind Cursor Composer 2. Cursor applied continued pretraining and high-compute RL training on top of it, with inference hosted via FireworksAI under an authorized commercial partnership.
KimiMar 18Kimi Team published Attention Residuals (AttnRes), replacing standard fixed residual connections in transformers with learned attention over preceding layer outputs. Block AttnRes matches a baseline trained with 1.25x more compute, improving across all evaluated benchmarks.
Frequently asked questions