{"id":6131,"date":"2025-04-14T06:46:36","date_gmt":"2025-04-14T06:46:36","guid":{"rendered":"https:\/\/www.myshirtai.com\/?p=6131"},"modified":"2025-04-14T06:46:36","modified_gmt":"2025-04-14T06:46:36","slug":"kimi-vl-a3b-%e5%8f%91%e5%b8%83%ef%bc%9a%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%8c128k%e4%b8%8a%e4%b8%8b%e6%96%87%e7%aa%97%e5%8f%a3%e4%b8%8emit%e8%ae%b8%e5%8f%af","status":"publish","type":"post","link":"https:\/\/www.myshirtai.com\/en\/archives\/6131","title":{"rendered":"Kimi VL A3B Released: Multimodal Large Model, 128K Context Window & MIT License"},"content":{"rendered":"

I. Introduction<\/strong><\/h3>\n

Recently.Moonshot AI has officially unveiled its latest generation of multimodal macromodeling<\/span>\u00a0<\/span>Kimi VL A3B<\/span>This is a lightweight model based on the Mixed Expert (MoE) architecture with 16B total parameters but only 2.8B activations for inference. its core highlights include<\/span>\u00a0<\/span>128K Extra Long Context Window<\/span><\/strong>,<\/span>multimodal reasoning ability<\/span><\/strong>. More excitingly, the model is in looseOpen under MIT license<\/strong>This not only highlights its technological breakthrough, but also provides unlimited possibilities for research and application. This paper will delve into the core features of Kimi VL A3B and its potential value.<\/p>\n

\"\"<\/p>\n

<\/h3>\n

II. Technical highlights: small models, big capabilities<\/strong><\/h3>\n

1. MoE architecture and lightweight design<\/h4>\n

The Kimi VL A3B utilizes the Mixed Expert (MoE) architecture, which significantly improves computational efficiency by dynamically assigning tasks to different expert sub-networks. Despite a total parameter of 16B, only 2.8B is activated during inference, allowing it to significantly reduce memory footprint and inference costs while maintaining performance. For example, in the MathVista Mathematical Reasoning Benchmark, KimiVL A3B achieves an accuracy of 68.7% with 2.8B active parameters, outperforming GPT-4o (68.5%) with a much larger parameter size.<\/p>\n

\"\"<\/p>\n

2. 128K context window, a new benchmark for long text processing<\/strong><\/h4>\n

Supporting a 128K context window, the Kimi VL A3B is capable of handling documents of tens of thousands of words, complex conversations, or multi-round interactive tasks. This feature enables it to excel in scenarios such as legal file analysis, technical document interpretation, and financial report generation. For example, in the MMLongBench-Doc long document comprehension test, Kimi VL A3B scored 35.1%, which is ahead of similar models.<\/p>\n

\"\"<\/p>\n

3. Multimodal capabilities: deep fusion of text, images and video<\/strong><\/h4>\n
    \n
  • \n
      \n
    • Visual Understanding: The native resolution visual encoder MoonViT supports high-resolution image input to parse complex diagrams, math equations and handwritten content without the need for slicing. It scored 867 in the OCRBench benchmark, achieving SOTA.<\/li>\n
    • Video Analytics: Ability to capture key details from hour-long video lessons and generate structured summaries.<\/li>\n
    • Cross-modal reasoning: Combine text and image information to solve geometry problems, analyze financial tables, and generate LaTeX code or Markdown tables.<\/li>\n
    • Comparison of Image Recognition Ability (Kimi-VL-A3B vs. GPT-4o): The content of the image is a screenshot from Cyberpunk 2077, and both of them are correct in parsing the content of the image, with GPT-4o parsing it faster, while Kimi-VL-A3B's answer is more comprehensive.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n