llm

Kimi K3 🌙 (Moonshot AI) Review (2026)

The world's first open 2.8T-parameter model with Kimi Delta Attention and Stable LatentMoE.

★ 4.9

Editorial Rating

Pricing Model:

Open Weights (Kimi License)

Recommended For:

Massive context reasoning (1M+ tokens), multimodal CAD/game engineering, and tool orchestration

Test Kimi K3 on Moonshot Web ↗

Technical Overview

Kimi K3 by Moonshot AI is an open-weight 2.8-trillion parameter multimodal titan. By scaling MoE sparsity to 896 experts and activating only 16 per token, Kimi K3 maintains lightning-fast execution while retaining world-class contextual fidelity across 1,000,000 tokens.

Quick Start Command

BASH / TERMINAL

pip install transformers compressed-tensors && vllm serve moonshotai/Kimi-K3

Advantages (Pros)

✓ 2.8T total parameters with Stable LatentMoE activating only 16 out of 896 specialized experts
✓ Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) yields 2.5× scaling efficiency over K2
✓ #1 among open-weights models on Toolathlon Verified (76.5%) and Terminal Bench 2.1 (88.3%)
✓ Native multimodal vision encoder supports vision-in-the-loop debugging and UI/CAD generation

Considerations (Cons)

✗ Requires specialized compressed-tensors library or custom vLLM fork for inference
✗ High aggregate memory bandwidth required to maintain high expert-selection throughput