Qwen3.8-Flash-Next
by Qwen
A 180B-parameter vision-language model: a describer for your pipeline, not an embedder
Qwen/Qwen3.8-Flash-NextDeploy Qwen3.8-Flash-Next
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Qwen3.8-Flash-Next is a mixture-of-experts vision-language model at 180 billion parameters, released in August 2026 as the preview of the architecture behind Qwen4. It takes images and video frames alongside text and produces text, which makes it a captioner, a visual question answerer and a structured-extraction model over pixels.
It produces no retrieval vector. That distinction decides where it belongs in a pipeline: a VLM describes a frame, and a text embedding model turns that description into something searchable. Teams reach for a model like this expecting an embedding and find an essay.
At 180B parameters with a sparse expert architecture, self-hosting is a cluster decision rather than a sidecar one. Alibaba's own card points at a hosted service for production use, and the weights are published under a community license rather than Apache-2.0, so read the terms before building a product on it.
Architecture
Mixture-of-experts decoder, 179,999,981,459 parameters total, architecture class Qwen4ExpForConditionalGeneration. Vision tower: 27 layers, 1152-dimensional hidden state, patch size 16, spatial merge 2 and temporal patch 2, so consecutive video frames are merged before they reach the language model, projecting to a 2560-dimensional model hidden state. Published for transformers, vLLM and SGLang.
Mixpeek SDK Integration
// A VLM produces TEXT. The pipeline shape is: describe the frame, embed the
// description, index both. The description is worth storing as a payload field
// as well, because it is what a reader sees in a result.
const res = await fetch(
"https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
{
method: "POST",
headers: {
Authorization: "Bearer API_KEY",
"Content-Type": "application/json",
},
body: JSON.stringify({
collection_id: "col_your_collection",
documents: [
{
document_id: "scene-00412-0184",
// Vector from a text embedding model over the generated description,
// not from the VLM.
vectors: { "text-embedding": yourDescriptionVector },
payload: {
source_key: "archive/2026/ep-412.mp4",
start_ms: 184000,
description: "Two people unpack a box on a kitchen counter; a price card reads 24.99",
},
},
],
}),
},
);Capabilities
- Image and video frame understanding with text output
- Temporal patching, so short frame sequences are processed together rather than one frame at a time
- Visual question answering and structured extraction from a frame
- Long-context text alongside the visual input
Use Cases on Mixpeek
Performance
Per-frame cost for a 180B mixture-of-experts model is set by your serving stack, and we have not measured it. Treat description generation as the expensive step in any pipeline that uses it, and sample rather than describing every frame.
Common Pipeline Companions
Frequently Asked Questions
Can I use Qwen3.8-Flash-Next as an embedding model?
No. It generates text from images and text, and it returns no retrieval vector. Use it to produce descriptions or structured fields, then embed that output with a text embedding model. Hidden states from a generative model can be pooled into something vector-shaped, and it is not trained for retrieval, so the result underperforms a purpose-built encoder.
How large is it really, and can I self-host it?
180 billion parameters in a mixture-of-experts layout, so only a fraction of the weights are active per token while all of them must be resident. That is a multi-GPU serving decision. Alibaba publishes a hosted variant for production use, and the card names it.
What license is it under?
qwen-community-1.0, a community license rather than Apache-2.0 or MIT. The terms are in the repo's LICENSE file, and they are worth reading before a commercial deployment.
Does Mixpeek run this model?
Not on the managed tier. A 180B VLM sits behind an inference endpoint you control, and the text it produces enters Mixpeek as a document field. On a single-tenant Enterprise deployment a custom plugin can call it during extraction.
Specification
Build a pipeline with Qwen3.8-Flash-Next
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free