Keye-VL-2.0-30B-A3B - Kuaishou's Open-Source Self-Developed Multimodal Large Model

What is Keye-VL-2.0-30B-A3B

Keye-VL-2.0-30B-A3B is Kuaishou's open-source self-developed multimodal large model, serving as the 30B-class primary foundation. The model is the first to introduce DSA sparse attention into multimodal scenarios, supports a 256K ultra-long context, and achieves millisecond-level temporal reasoning over hour-long videos. It surpasses Gemini-2.5-Pro and Gemini 3 Flash on the TimeLens benchmark, and is the first to unlock Agent collaboration mechanisms such as Code, Tool, and Search, evolving the model from an observer into an actor.

Key Features of Keye-VL-2.0-30B-A3B

• Ultra-Long Video Understanding: Supports a 256K ultra-long context, capable of processing hour-long video sequences and achieving near-lossless deep temporal reasoning.
• Temporal Causal Reasoning: Captures the causal chain behind the visuals within a continuous temporal flow, achieving the leap from "seeing the visuals" to "understanding the logic".
• Millisecond-Level Frame-Level Localization: Possesses scalpel-like fine-grained parsing capability, enabling precise timestamp-level breakdown of complex process flows or gaming highlights.
• Deep Cross-Modal Fusion: Simultaneously processes visual, audio, and textual information, achieving collaborative understanding and deep semantic alignment across modalities.
• Agent Collaborative Execution: The first to unlock system-level autonomous collaboration and task execution capabilities in complex scenarios such as code generation, tool invocation, and search.
• High-Noise Information Purification: Accurately captures key frames and clarifies dynamic patterns in complex scenarios, effectively filtering out redundant information and retaining core content.

Technical Principles of Keye-VL-2.0-30B-A3B

• DSA Sparse Attention Mechanism: The first to introduce DeepSeek Sparse Attention into multimodal understanding, combining sparse attention with targeted feature aggregation to break through the exponential computational bottleneck of ultra-long visual contexts.
• Ultra-Long Context Architecture: Adopts a 256K Token-level end-to-end architecture, achieving coherent deep perception of long video sequences without segmentation or truncation processing.
• Fine-Grained Temporal Understanding Engine: Through frame-level action boundary recognition, dynamic visual parsing, and audio-visual collaborative modeling, achieves millisecond-level precise temporal localization and causal inference.
• Agent Collaboration Framework: Integrates Code Interpreter, Tool Use, and Search capabilities, building a closed-loop decision system from multimodal perception to logical reasoning to tool execution.
• Unified Multimodal Feature Fusion: Maps visual, audio, and textual features into a shared representation space, achieving deep semantic alignment and joint reasoning of cross-modal information.

How to Use Keye-VL-2.0-30B-A3B

• Obtain the Model: Fully open-source model weights and deployment documentation can be downloaded via GitHub, Hugging Face, or ModelScope.
• Hardware Preparation: Requires H800 or equivalent computing power GPUs, and the use of at least two GPUs for multi-GPU tensor parallel inference.
• Docker Quick Deployment: Directly pull the official Docker image and run it to complete environment configuration and model loading with a single click.
• Source Code Installation and Deployment: Sequentially clone the three dependency repositories—Keye-customized SGLang, DeepGEMM, and EffectiveKernels—and complete the compilation and installation.
• Start the Inference Service: Load the model weights with SGLang, set the tensor parallelism parameters and enable remote code trust, and you can launch a locally hosted API service compatible with the OpenAI protocol.
• Call the API: After launch, send video and text instructions via standard HTTP requests, and the model will return structured long-video understanding results or Agent execution output.

Core advantages of Keye-VL-2.0-30B-A3B

• DSA's first deployment in multimodal: The first introduction of DeepSeek Sparse Attention into multimodal understanding scenarios, fundamentally solving the exponential computing bottleneck brought by ultra-long visual contexts and enabling efficient inference on hour-level videos.
• 256K ultra-long context: Supports a token-level ultra-long context of up to 256K, enabling nearly lossless end-to-end deep perception of hour-level video sequences without the segmented truncation processing required by traditional models.
• Millisecond-level frame-level localization: With scalpel-like fine-grained temporal parsing capability, it can precisely decompose and localize every key action in scenarios such as complex industrial processes and gaming highlights down to the timestamp.
• Temporal causal reasoning: Going beyond simple frame label recognition, it captures causal chains in continuous temporal flows, achieving a leap from "seeing the picture" to "understanding the logic"; for example, it can directly infer from a "snowy car accident" scene the safety strategy that "joining a tour group is better than self-driving."
• Agent collaboration mechanism: For the first time, the Keye series unlocks system-level autonomous collaboration and execution capabilities in complex scenarios such as Code, Tool, and Search, evolving the model from a passive "observer" into an active "actor" that solves tasks.