AI Research Atlas

Kimi-VL

Moonshot AI · 10 April 2025

Open MoE vision-language model activating only 2.8B parameters (16B total) with 128K context and agent capabilities such as OSWorld; Thinking variant followed.

Small, efficient open VLM (MoonViT encoder plus an MoE language decoder) aimed at long-context video, OCR and GUI-agent tasks. A reasoning variant, Kimi-VL-Thinking, followed in mid-2025.

Date
Thursday, 10 April 2025
Lab
Moonshot AI
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
Active / total parameters2.8B / 16B
language decoder		company

MIT license.

Sources

  1. arxiv.org/abs/2504.07491
  2. huggingface.co/moonshotai/Kimi-VL-A3B-Instruct

This record was checked against its sources on 6 October 2026. How we check