Kimi K2.5

• 🧠 Trillion-parameter Mixture-of-Experts, 32B active per token.
• 🆕 Native multimodal: continual pretraining on ~15T mixed vision-text tokens.
• 📏 256K-token context window for long documents and codebases.
• 👁️ Adds vision and cross-modal reasoning over text-only Kimi K2.
• 🔧 Instant and thinking modes for speed or deeper reasoning.
• ⚡ Native INT4 quantization for efficient inference.
• 🎯 Provider-reported HLE with tools: 51.8 text, 39.8 image.
• 🌐 Open-weight with OpenAI/Anthropic-compatible API.

Moonshot

Moonshot is an AI research lab known for developing the Kimi family of large language models. The organization has gained recognition for building capable reasoning-oriented models, with the Kimi line representing its flagship series of text generation…

Kimi K2.5, released January 2026 by Beijing-based Moonshot AI, is the multimodal evolution of the Kimi K2 line. It keeps the family's trillion-parameter Mixture-of-Experts design — with roughly 32 billion active parameters per token and a 256K context window — while adding a vision pathway. Per Moonshot's model card, K2.5 was built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base, so vision and language develop together rather than as bolted-on features.

The clearest gain over the text-only predecessor Kimi K2 is native multimodality: K2.5 understands images, supports cross-modal reasoning, and grounds agentic tool use in visual inputs. It also formalizes dual operation — an instant mode for speed and a thinking mode for deeper reasoning. On Moonshot's reported Humanity's Last Exam, K2.5 scores 31.5 text and 21.3 image without tools, rising to 51.8 text and 39.8 image with tools.

K2.5 became the architectural template for its successors. Kimi K2.6 reuses the same trillion-parameter, 32B-active topology, changing the post-training recipe, while Kimi K2.7 Code specializes the same MoE backbone for long-horizon software engineering. All ship open-weight with native INT4 quantization and an OpenAI/Anthropic-compatible API.