Kimi-VL-A3B-Thinking
Active Parameters
16B
Context Length
128K
Modality
Multimodal
Architecture
Mixture of Experts (MoE)
License
MIT License
Release Date
10 Apr 2025
Knowledge Cutoff
Oct 2024
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-Thinking available.
Rankings
Overall Rank
-
Coding Rank
-
About Kimi-VL-A3B-Thinking
Kimi-VL-A3B-Thinking is an advanced vision-language model (VLM) developed by Moonshot AI, engineered to bridge the gap between efficient parameter utilization and high-fidelity multimodal reasoning. Architecturally, it is built upon the Mixture-of-Experts (MoE) framework of the Moonlight LLM series, integrating a proprietary native-resolution visual encoder known as MoonViT via an MLP projector. The model is specifically optimized for long-horizon cognitive tasks through supervised fine-tuning and reinforcement learning, allowing it to generate extended chains of thought (CoT) when processing complex visual and textual inputs.
The system utilizes a sparse MoE design comprising 16 billion total parameters, with only approximately 2.8 billion parameters activated during any single inference step. The language decoder follows a configuration similar to the DeepSeek-V3 architecture, featuring Multi-head Latent Attention (MLA) and a specialized gating mechanism that routes tokens through 64 routed experts. This structural innovation enables the model to handle diverse input resolutions and aspect ratios without downsampling, preserving the fidelity of visual data for tasks such as optical character recognition (OCR) and college-level academic analysis.
Functionally, Kimi-VL-A3B-Thinking supports an expansive context window of 128,000 tokens, facilitating the ingestion of lengthy documents, multi-image sequences, and video content. The "Thinking" variant is tailored for scenarios requiring multi-step mathematical problem-solving, document comprehension, and autonomous agent interactions. By leveraging Flash-Attention 2 and supporting native half-precision formats, the model maintains high throughput and computational efficiency across a broad spectrum of multimodal reasoning applications.
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
27
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
64
Active Experts
2
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.