Qwen3.8-27B

by Qwen

Open-weight vision-language model for reading images and video frames as text

1.0Mdl/month

27Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

Qwen/Qwen3.8-27B

Feature URI

mixpeek://image_extractor@v1/qwen38_27b_v1

Overview

Qwen3.8-27B is the vision-language generation in the Qwen3.x line, taking interleaved image and text input and producing text. It is the most-liked image-text-to-text model on the Hugging Face Hub at the time of writing, with roughly 1.0M downloads, and it ships under Apache 2.0, which makes it one of the few frontier-scale VLMs you can self-host without a negotiated licence.

On Mixpeek it is used as a captioning and visual-question-answering stage over extracted frames: scene-split a video, sample keyframes, and have the model describe what is in them so the text is searchable alongside the visual embedding.

Architecture

Decoder-only vision-language model, 27B parameters, qwen3_5 architecture in transformers. Conversational, accepts interleaved image and text input. Also published as an FP8 checkpoint for reduced memory footprint.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mixpeek = new Mixpeek({ apiKey: process.env.MIXPEEK_API_KEY });

// Caption sampled frames so the description is searchable next to the embedding.
await mixpeek.collections.create({
  collection_name: "video_frames",
  feature_extractors: [
    { extractor: "mixpeek://image_extractor@v1/qwen38_27b_v1" },
  ],
});