Moonshot AI's open-weight multimodal reasoning model at very large scale. Built for complex coding, knowledge work and long-horizon agent runs, with particular strength in navigating large repositories and iterating against images, logs, tests and runtime feedback. Uses KDA and Attention Residuals for efficiency.

The coding model in Moonshot AI's K2.7 line, built to finish programming tasks end to end over long contexts. A native multimodal mixture of experts taking text, image and video, always in thinking mode, with the full reasoning kept across turns. 256K context, 32B active parameters out of roughly 1T.

Moonshot AI's K2.6, aimed at long-horizon coding, UI generation from prompts or images, and multi-agent orchestration. It works end to end across Python, Rust and Go, and its agent swarm scales to hundreds of parallel sub-agents, returning documents, websites and spreadsheets in one run without supervision.

Moonshot AI's natively multimodal K2.5, continuing pretraining on Kimi K2 over roughly 15T mixed visual and text tokens. Its two selling points are visual coding and a self-directed agent swarm, alongside general reasoning and tool calling.

Kimi Linear is a hybrid linear attention architecture that outperforms traditional full attention methods. Features Kimi Delta Attention (KDA) for efficient memory usage, reducing KV caches by up to 75% and boosting throughput by up to 6x for contexts as long as 1M tokens.

The reasoning model in Moonshot AI's K2 line, on the same trillion-parameter mixture-of-experts backbone with 32 billion active per pass and 256K context. It interleaves thinking with tool calls, and is built to hold a research or coding task together across hundreds of sequential actions without drifting.

The September refresh of Kimi K2, still 1 trillion total parameters with 32 billion active per pass, now holding 256K context instead of 128K. The update targets agentic coding accuracy across different scaffolds and front-end output worth looking at, for web and 3D work.

Moonshot AI's Kimi K2, a mixture-of-experts model with 1 trillion total parameters and 32 billion active per forward pass. Tuned for agentic work: tool use, reasoning and code synthesis, with 128K context. Trained with the MuonClip optimizer, which is what kept a MoE run at this scale stable.

Moonshot AI's open-weight multimodal reasoning model at very large scale. Built for complex coding, knowledge work and long-horizon agent runs, with particular strength in navigating large repositories and iterating against images, logs, tests and runtime feedback. Uses KDA and Attention Residuals for efficiency.

The coding model in Moonshot AI's K2.7 line, built to finish programming tasks end to end over long contexts. A native multimodal mixture of experts taking text, image and video, always in thinking mode, with the full reasoning kept across turns. 256K context, 32B active parameters out of roughly 1T.

Moonshot AI's K2.6, aimed at long-horizon coding, UI generation from prompts or images, and multi-agent orchestration. It works end to end across Python, Rust and Go, and its agent swarm scales to hundreds of parallel sub-agents, returning documents, websites and spreadsheets in one run without supervision.

Moonshot AI's natively multimodal K2.5, continuing pretraining on Kimi K2 over roughly 15T mixed visual and text tokens. Its two selling points are visual coding and a self-directed agent swarm, alongside general reasoning and tool calling.

Kimi Linear is a hybrid linear attention architecture that outperforms traditional full attention methods. Features Kimi Delta Attention (KDA) for efficient memory usage, reducing KV caches by up to 75% and boosting throughput by up to 6x for contexts as long as 1M tokens.

The reasoning model in Moonshot AI's K2 line, on the same trillion-parameter mixture-of-experts backbone with 32 billion active per pass and 256K context. It interleaves thinking with tool calls, and is built to hold a research or coding task together across hundreds of sequential actions without drifting.

The September refresh of Kimi K2, still 1 trillion total parameters with 32 billion active per pass, now holding 256K context instead of 128K. The update targets agentic coding accuracy across different scaffolds and front-end output worth looking at, for web and 3D work.

Moonshot AI's Kimi K2, a mixture-of-experts model with 1 trillion total parameters and 32 billion active per forward pass. Tuned for agentic work: tool use, reasoning and code synthesis, with 128K context. Trained with the MuonClip optimizer, which is what kept a MoE run at this scale stable.