Kimi K3

Kimi K3 is Moonshot AI’s flagship open-weights frontier model engineered for long-horizon software engineering, complex reasoning, and multimodal agentic workflows. As a 3-trillion-parameter-class sparse Mixture-of-Experts (MoE) model, Kimi K3 delivers frontier-level performance with variable reasoning effort and native visual understanding.

Model Overview

• Developer: Moonshot AI
• Model Class: Frontier Open-Weight Agentic Engine
• Total Parameters: 2.8 Trillion
• Active Parameters: 104 Billion (16 / 896 experts activated per token)
• Architecture: Sparse Mixture-of-Experts (MoE) with Kimi Delta Attention (KDA) & Attention Residuals (AttnRes)
• Context Length: 1,048,576 tokens (1M Context Window)
• Vision Encoder: MoonViT-V2 (401M parameters)
• Native Quantization: MXFP4 weights / MXFP8 activations
• License: Kimi K3 Open Weights License

Core Capabilities & Performance Profile

Long-Horizon Agentic Coding

Operates across extended development sessions with high autonomy—navigating large repositories, compiling system code, optimizing low-level GPU kernels, and resolving distributed race conditions without losing contextual coherence.

Hybrid Linear Attention (KDA)

Combines standard transformer mechanics with Kimi Delta Attention across 93 layers. This hybrid approach preserves long-range sequence coherence across its 1M context window without incurring quadratic throughput degradation.

Vision-in-the-Loop Multi-Modality

Integrates MoonViT-V2 to parse complex visual artifacts directly alongside code. It processes architectural diagrams, UI mockups, terminal screenshots, and CAD assets within the same logical context window.

Dynamic Thinking Mode

Features built-in reasoning traces enabled by default. Supports dynamic inference depth via the reasoning_effort flag (low, high, max), allowing latency tuning based on problem difficulty.

Technical Specifications

Resource Requirements & Quantization

Running Kimi K3 locally requires enterprise GPU clusters or high-bandwidth unified memory systems: