Kimi-VL (A3B Instruct and Thinking)

2025-04-10 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI releases Kimi-VL, a 16-billion-parameter MoE vision-language model (2.8 billion active in the decoder), with a reasoning variant, under the MIT license and with a 128K context.

### Context In spring 2025, the most capable open vision-language models (Qwen2.5-VL 72B, Gemma 3) are dense and heavy to deploy. Moonshot AI offers a lightweight MoE alternative, in line with the path initiated by Kimi k1.5 (multimodal reasoning). An improved version, Kimi-VL-A3B-Thinking-2506, will follow in June 2025: it is grouped into this entry. ### What the model brings - MoonViT native-resolution vision encoder, which processes very high-resolution images without splitting them into fixed-size tiles. - MoE decoder with only 2.8 billion active parameters, for a reduced inference cost. - 128K-token context for long documents and videos (64.5 on LongVideoBench and 35.1 on MMLongBench-Doc, according to the card). - Thinking variant obtained through long chain-of-thought fine-tuning followed by RL, with 61.7 on MMMU, 36.8 on MathVision and 71.3 on MathVista. - 2506 version: 56.9 on MathVision, 80.1 on MathVista, images up to 3.2 million pixels and about 20% less thinking (according to the vendor). ### Architecture and specifications 16 billion parameters in total and 3 billion activated (2.8 B in the decoder), MoonViT encoder, 128K context. Variants: Kimi-VL-A3B-Instruct and Kimi-VL-A3B-Thinking (then Thinking-2506). License: MIT. Weights: huggingface.co/moonshotai/Kimi-VL-A3B-Thinking. Report: arXiv 2504.07491. ### Cost and access | Item | Detail | |---|---| | Official API at launch | Not disclosed in the sources consulted | | Self-hosting | 16 B model in total; required hardware not disclosed | | License | MIT | ### Benchmarks Thinking variant, scores published by the vendor (model card), with closed models as a reference. | Benchmark (Pass@1) | GPT-4o | OpenAI o1-1217 | Kimi-k1.5 | Kimi-VL-Thinking | |---|---|---|---|---| | MathVision (full) | 30.4 | not disclosed | 38.6 | 36.8 | | MathVista (mini) | 63.8 | 71.0 | 74.9 | 71.3 | | MMMU (val) | 69.1 | 77.3 | 70.0 | 61.7 | Source: Hugging Face card of Kimi-VL-A3B-Thinking, vendor measurements. ### Versus closed models On MathVision and MathVista, the Thinking variant outperforms GPT-4o according to the vendor's figures (36.8 versus 30.4; 71.3 versus 63.8), and comes close to o1-1217 on MathVista (71.3 versus 71.0). It remains behind on MMMU (61.7 versus 69.1 for GPT-4o and 77.3 for o1-1217), a general-knowledge benchmark where its active size weighs in. As no API pricing is disclosed, the cost gap cannot be quantified. Its benefit for the enterprise is a self-hostable VLM under the MIT license, useful when images must not leave the infrastructure. ### Key takeaways Kimi-VL is the family's open MoE VLM, lightweight and under the MIT license. It shows that multimodal reasoning can fit in a model with about 3 billion active parameters, with limits on general knowledge. The two variants (Instruct and Thinking) make it possible to choose between fast perception and long reasoning.

Tags : LLM, Vision, Open Source, Moonshot AI, Kimi, MIT, MoE