Kimi K2

Frontier open-weight Mixture-of-Experts large language model from Moonshot AI. Architecture: **1T total parameters, 32B activated per token**, 384 experts (8 selected + 1 shared per layer), 61 layers, and Multi-head Latent Attention (MLA) for KV-cache compression. Released initially in 2025; the current K2.6 release (April 20, 2026) adds a 400M-parameter MoonViT vision encoder for native multimodal input, 256K context across all variants, and an Agent Swarm system that scales to 300 sub-agents and 4,000 coordinated steps per query. Modified MIT license — weights are downloadable.

The 2025-2026 frontier of open-weight models is dominated by aggressively-sparse MoE architectures — getting frontier-class capability at fraction-of-dense inference cost. Kimi K2 is Moonshot AI's bet that a tightly-routed 32-of-1000B activation pattern beats dense alternatives on both training efficiency and serving cost. The 2026 evolution (K2.6) extends that bet to long-context agentic workloads where the **disaggregated prefill** architecture published in Moonshot's open-source Mooncake serving platform makes million-token-context agent runs economically viable.

Agentic coding assistants (Kimi Code Review claims 88% coding-cost reduction vs hosted frontier alternatives), high-context-window RAG over corpus documents (256K context), multimodal document analysis via MoonViT, large-scale agent-swarm coordination (up to 300 sub-agents), and any open-weight deployment where frontier-class quality + MIT license matter (sovereign clouds, regulated industries).