Keye-VL-2.0-30B-A3B

by Kwai-Keye

Kuaishou's video-centric vision-language model for clip understanding and Q&A

290dl/month

30B MoE (~3B active)params

HuggingFace Run on your data, free

Identifiers

Model ID

Kwai-Keye/Keye-VL-2.0-30B-A3B

Feature URI

mixpeek://video_extractor@v1/kwai_keye_vl2_30b_a3b_v1

Overview

Keye-VL 2.0 (30B Mixture-of-Experts with ~3B active params) is Kuaishou's video-first multimodal LLM, built for understanding short-form and long video alongside images and text. It is strong at video question answering, captioning, and temporal reasoning over clips.

On Mixpeek, a model like Keye-VL works as a captioning/understanding stage in a video pipeline: a fast encoder (V-JEPA 2, VideoPrism, InternVideo2) retrieves candidate clips, then a VLM like Keye-VL generates grounded descriptions or answers questions about the retrieved moments — keeping the expensive VLM off the full corpus.

Architecture

Mixture-of-Experts vision-language model (~30B total, ~3B active) with a vision encoder feeding an LLM decoder, instruction-tuned for video and image understanding, captioning, and VQA with temporal reasoning.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

// Managed: create a collection over a bucket; Mixpeek runs this model's extractor
const collection = await mx.collections.create({
  namespace_id: "my-namespace",
  collection_name: "my-collection",
  source: { type: "bucket", bucket_ids: ["bkt_your_bucket"] },
  feature_extractor: {
    feature_extractor_name: "video_description",
    version: "v1",
    parameters: { model_id: "Kwai-Keye/Keye-VL-2.0-30B-A3B" },
  },
});