Kimi-VL

What it is

Kimi-VL is an open-source vision-language model released by Moonshot AI in April 2025 ^[1]. A vision-language model takes an image together with text and answers in text: describe a chart, read a screenshot, solve a problem written on a photographed page.

Architecturally it is a mixture-of-experts (MoE) model with 16B parameters in total and roughly 3B active ^[1]. Mixture-of-experts is a way of building a large network out of many smaller sub-networks, called experts, plus a routing step that sends each token to only a few of them. Every parameter is stored, but only a fraction is used for any given token, so the model runs at the cost of a much smaller dense model. Kimi-VL's model card puts 2.8B of the active parameters in the language decoder and about 0.4B in the MoonViT vision tower ^[3].

Releases and versions

Two versions, Kimi-VL-A3B-Instruct and Kimi-VL-A3B-Thinking, were posted to Hugging Face on 9 April 2025, and the MoonViT-SO-400M vision encoder followed on 10 April ^[2]. The split follows a now-common pattern: an instruct model that answers directly, and a thinking model that works through intermediate reasoning before answering. An updated thinking model, Kimi-VL-A3B-Thinking-2506, appeared in June 2025 ^[1][2].

Design choices

The model card describes MoonViT as a native-resolution vision encoder that handles ultra-high-resolution inputs ^[3]. Instead of squashing every image into one fixed small size, a native-resolution encoder keeps the picture closer to the size it arrived at, which matters when the content is dense — a table, a document scan, a screen full of small interface elements. The card also gives a 128K extended context window and an MIT license ^[3], so the weights can be used and modified commercially with few strings attached.

Reported results

The model card reports the following scores ^[3]:

• MMMU-Val 57.0, a multi-discipline benchmark of college-level questions that require images.
• MMBench-EN-v1.1 83.1, a broad English evaluation of visual understanding.
• MathVista 68.7, which tests mathematical reasoning when the problem is given visually.
• OCRBench 867, which measures reading text out of images.
• ScreenSpot-V2 92.8 and ScreenSpot-Pro 34.5, both testing whether a model can locate a named element on a screen — the kind of ability an interface-controlling agent needs.

The gap between the two ScreenSpot figures is the striking part: locating elements on ordinary screens looks close to solved at 92.8, while the harder Pro setting sits at 34.5.

Where it sits

Kimi-VL belongs to a cluster of open releases Moonshot AI put out in the same period. Moonlight-16B-A3B came in February 2025, Kimi-Audio-7B in April 2025, and Kimi-Dev-72B — a coding model based on Qwen2.5-72B — in June 2025 ^[1][2]. The "16B-A3B" naming in Kimi-VL echoes Moonlight and signals the same economy: a large pool of parameters with a small active slice per token.

The vision side proved reusable. The roughly 400M-parameter MoonViT encoder was later used in Kimi K2.5 in January 2026 ^[1], which makes Kimi-VL not just a standalone model but the point where Moonshot's vision component was introduced and tested in the open.