#4Moonshot AI (Kimi)

Kimi K3

Moonshot's flagship and, by its own description, the first open model in the 3-trillion-parameter class. It handles text, images and video in one model and targets long coding and knowledge-work tasks. Moonshot says it still trails the strongest closed models from Anthropic and OpenAI.

• 2.8T total / 104B active MoE (16 of 896 experts)
• 1,048,576-token context
• Text, image and video input
• Kimi Delta Attention; native MXFP4 weights
• Thinking levels: low, high, max

Accesskimi.ai web and apps, Kimi Work, Kimi Code and Kimi API from launch; open weights on Hugging Face from 2026-07-27 under the Kimi K3 License

Price$3.00 input / $15.00 output per 1M tokens; cached input $0.30

Kimi K2.7 Code

A coding-focused open model built on Kimi K2.6 for long, real-world programming tasks. Moonshot says it uses about 30% fewer reasoning tokens than K2.6, which makes automated coding cheaper, and it became the model behind Moonshot's Kimi For Coding service. A faster, higher-priced 'highspeed' API tier is also offered.

• 1T total / 32B active MoE
• 256K context (262,144 tokens on the API)
• Image and video input via MoonViT encoder
• About 30% fewer reasoning tokens than K2.6 (lab claim)

AccessKimi Code (Kimi For Coding); API (kimi-k2.7-code and a highspeed tier); open weights on Hugging Face, Modified MIT licence

Price$0.95 input / $4.00 output per 1M tokens; cache hit $0.19

Kimi K2.6

An open multimodal agent model aimed at long coding jobs, coding-driven design and coordinating swarms of sub-agents. Moonshot says it can scale to 300 sub-agents running 4,000 coordinated steps. It keeps the 1-trillion-parameter K2 design and remains on Moonshot's API price list alongside K3.

• 1T total / 32B active MoE
• 256K context
• Text, image and video input
• SWE-Bench Verified 80.2 (lab model card)
• Agent Swarm: up to 300 sub-agents (lab claim)

AccessKimi app and API (kimi-k2.6); open weights on Hugging Face, Modified MIT licence

Price$0.95 input / $4.00 output per 1M tokens; cache hit $0.16

Kimi K2.5

Moonshot's first natively multimodal K2 model, combining image and video understanding with tool use and visual code generation. It introduced 'Agent Swarm', which splits a large task across many parallel sub-agents. Weights were published under the same Modified MIT licence as earlier K2 models.

• 1T total / 32B active MoE
• 256K context
• Text, image and video input (MoonViT 400M encoder)
• SWE-Bench Verified 76.8, thinking mode (lab model card)

AccessKimi app and API; open weights on Hugging Face, Modified MIT licence

Kimi K2 Thinking

A reasoning version of Kimi K2 trained to think step by step while calling tools. Moonshot says it stays on task across 200 to 300 tool calls in a row and runs natively in INT4 for roughly twice the generation speed.

• 1T total / 32B active MoE
• 256K context
• Native INT4 inference (quantization-aware training)
• HLE with tools 44.9; BrowseComp 60.2 (lab model card)

AccessKimi app and API; open weights on Hugging Face, Modified MIT licence

Kimi Linear (48B-A3B)

A small open research model testing Kimi Delta Attention, a linear-attention design. Moonshot says it cuts the memory needed for long contexts by up to 75% and speeds up output up to six times at 1M tokens. The same attention later appears in Kimi K3.

• 48B total / 3B active
• 1M-token context
• KV cache up to 75% smaller (lab claim)
• Up to 6x faster per-token decoding at 1M context (lab claim)

AccessOpen weights on Hugging Face (Base and Instruct), MIT licence