Kimi-VL-A3B-Instruct
Active Parameters
16B
Context Length
128K
Modality
Multimodal
Architecture
Mixture of Experts (MoE)
License
MIT
Release Date
10 Apr 2025
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
2x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
3x RTX 4090
24GB VRAM
Datacenter
1x NVIDIA A100
80GB VRAM
Apple Silicon
1x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for Kimi-VL-A3B-Instruct available.
Rankings
Overall Rank
-
Coding Rank
-
About Kimi-VL-A3B-Instruct
Kimi-VL-A3B-Instruct is a multimodal Mixture-of-Experts (MoE) vision-language model developed by Moonshot AI, designed for high-resolution visual perception and long-context reasoning. The model operates on a base architecture that integrates a native-resolution visual encoder, termed MoonViT, with a sparse MoE language decoder. This design facilitates the processing of diverse inputs including single and multi-image sets, video sequences, and extensive document formats. The model is instruction-tuned to support interactive chat and agentic workflows, emphasizing efficiency in both high-resolution image analysis and natural language understanding across extended sequences.
Technically, the model utilizes a sparse MoE language backbone named Moonlight, which contains 16 billion total parameters but activates only 2.8 billion parameters per token. This sparsity is achieved through a routing mechanism that selects 8 experts from a total pool of 384 available experts. The visual component, MoonViT, supports native resolution processing up to 1792x1792 pixels, allowing the model to maintain high fidelity for OCR and detailed graphical analysis without forced resizing. The architecture incorporates a variable-length sequence attention mechanism that is compatible with FlashAttention, ensuring computational efficiency when handling images of various aspect ratios and resolutions.
Kimi-VL-A3B-Instruct is optimized for complex multimodal tasks such as document parsing, long-form video comprehension, and interactive GUI agent operations. Its large context window of 128,000 tokens enables the ingestion of multiple high-resolution images or lengthy video clips alongside extensive textual prompts. By combining the efficiency of MoE with high-resolution visual encoding, the model is suited for applications requiring detailed visual grounding and the ability to reason over long-form, multi-source information in a conversational or agentic context.
Technical Specifications
Attention
Attention Structure
Multi-Head Attention
Attention Heads
16
Key-Value Heads
16
Attention Head Dimension
-
Position Embedding
Absolute Position Embedding
RoPE Theta
800,000
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
2,048
Number of Layers
24
FFN Intermediate Size (Dense)
1,408
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
163,840
Mixture of Experts
Total Expert Parameters
3.0B
Number of Experts
384
Active Experts
8
Shared Experts
2
FFN Intermediate Size (per Expert)
1,408
Dense Layers Before MoE
1
About Kimi-VL
Kimi-VL by Moonshot AI is an efficient, open-source Mixture-of-Experts vision-language model. It employs a native-resolution MoonViT encoder and an MoE language model, activating 2.8 billion parameters. The model handles high-resolution visual inputs and processes contexts up to 128K tokens. A "Thinking" variant provides enhanced long-horizon reasoning.