Kimi K3

by Moonshot AI

Frontier flagship (July 2026). 2.8T parameters (MoE), 1M context, KDA hybrid attention, native multimodal, open weights (custom Kimi K3 License). Leaderboard ranks previously listed here (AA #3, Arena frontend #1) are unverified snapshots; check the live boards.

Release

• Date: July 16, 2026
• Open weights: Hugging Face moonshotai/Kimi-K3; press says 2026-07-26 or 07-27 (sources differ)
• License: custom “Kimi K3 License” (model card). Secondary reports say it gates commercial use above ~$20M annual revenue; the earlier “Modified MIT” label is not what the model card says. Read the license before commercial use.
• Parameters: 2.8T total / 104B active per the model card (earlier note said ~50B: corrected)

Architecture

Attention:

• KDA (Kimi Delta Attention): 3:1 hybrid linear-to-full attention
• Attention Residual (AttnRes) for information flow
• MXFP4 weights, MXFP8 activations

Scaling:

• 896 experts, 16 active per token; 93 layers (model card)
• 2.5× better compute-to-intelligence vs K2
• 6.3× faster long-context decoding

Capabilities

Context:

• 1 million tokens (~750K words)
• 64K max output
• 6.3× faster decoding than K2 for long-context

Multimodal:

• Native vision (MoonViT-V2 encoder per model card; text, image and video)
• Image + text at 1M context (unique)
• Charts, diagrams, documents

Reasoning:

• Always-on by default
• Multi-step decomposition
• 300 Agent Swarm (4K+ coordinated steps)

Benchmarks (unverified, vendor/press-reported)

• Artificial Analysis: Rank #3 overall
• Arena.ai Frontend Code: Rank #1
• Agentic: ~89.5% (5th frontier capability)
• Real-world: 5 of 6 wins vs Claude Sonnet 5

Pricing model

• Usage-based API pricing per token, with cheaper cache-hit input (unverified; check platform.kimi.ai)
• Cost advantage: 50-65% per-task vs Claude Sonnet 5

Availability

• Hosted: platform.kimi.ai, kimi.com
• Open weights: Hugging Face (July 27)
• Self-hosting: vLLM support announced

Positioning

• Frontier-class reasoning + agentic autonomy
• Open-weight advantage vs Claude/GPT
• Cost-efficient for complex tasks
• Chinese AI flagship release

See Also

• Kimi K2.6
• kimi
• Moonshot AI

Sources

• https://huggingface.co/moonshotai/Kimi-K3 (model card, 2026-09-30)
• https://qz.com/moonshot-ai-kimi-k3-open-weights-download-072726 (search snippet; 403 on fetch)
• https://roo.beehiiv.com/p/kimi-k3-open-weights-license-benchmarks (search snippet)