[Models](/models)/[Captioning](/models?category=captioning)/ParaVT/ParaVT-8B

HFScene CaptioningApache 2.0

# ParaVT-8B

by ParaVT

Agentic long-video model trained for parallel temporal tool calls

544dl/month

8B classparams

[HuggingFace](https://huggingface.co/ParaVT/ParaVT-8B)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`ParaVT/ParaVT-8B`

## Deploy ParaVT-8B

Single-tenant

Mixpeek 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.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

ParaVT-8B is a video-text-to-text model focused on long-video understanding through tool use. Its model card describes a parallel video tool-calling approach where the model can dispatch multiple temporal crop requests in one turn instead of walking sequentially through a video.

On Mixpeek, ParaVT is relevant when an agent needs to search long clips and then decide which time windows to inspect. It belongs after retrieval: use vector, transcript, or scene search to narrow the corpus, then let ParaVT reason over candidate spans and request focused temporal crops.

## Architecture

Final post-RL ParaVT checkpoint based on Qwen3VLForConditionalGeneration and Qwen/Qwen3-VL-8B-Instruct. The release uses cold-start SFT followed by PARA-GRPO reinforcement learning for parseable, parallel temporal tool-calling behavior.

## Mixpeek SDK Integration

IngestRetrieve

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so ParaVT-8B runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace\_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
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: \[
        {
          document\_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted\_text: modelOutput, source\_key: "archive/2026/asset-00412" },
          vectors: { "text-embedding": embeddingOfModelOutput },
        },
      \],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal\_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

## Capabilities

*   Video-text-to-text reasoning over long clips
*   Parallel temporal crop tool calls
*   Agentic RL training for video tool use
*   Apache 2.0 license
*   Drop-in Transformers and vLLM deployment pattern

## Use Cases on Mixpeek

Let agents inspect several candidate video windows in one reasoning pass

Find evidence moments inside long recordings after first-stage retrieval

Build video QA systems that use temporal crop tools

Compare sequential and parallel video inspection strategies

## Benchmarks

Dataset

Metric

Score

Source

ParaVT release

Task focus

Long-video parallel tool calling

ParaVT model card

Hugging Face

Release age

Created May 2026

HF model metadata

## Performance

Input SizeLong videos through temporal windows

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU Memory8B class VLM

Best used after retrieval has narrowed the video corpus or candidate time spans

## Common Pipeline Companions

[

facebook/vjepa2-vitl-fpc64-256

First-stage video embedding recall

](/model/facebook/vjepa2-vitl-fpc64-256)[

openai/whisper-large-v3

Transcript search for spoken clues

](/model/openai/whisper-large-v3)[

TencentARC/TimeLens-8B

Temporal grounding comparison for video evidence

](/model/TencentARC/TimeLens-8B)

## Explore on Mixpeek

[

More Captioning models

Compare alternatives in this category

](/models?category=captioning)[

Best Document AI Platforms

Hand-picked tools & platforms compared

](/curated-lists/best-document-ai-platforms)[

OCR & Document AI Internals

Deep-dive technical guide

](/guides/ocr-document-ai-internals)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

FrameworkHF

OrganizationParaVT

FeatureScene Captioning

Outputtext

Modalitiesvideo, image

RetrieverSemantic Search

Parameters8B class

LicenseApache 2.0

Downloads/mo544

[

### Research Paper

ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning

arxiv.org](https://arxiv.org/abs/2605.20342)

### Build a pipeline with ParaVT-8B

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](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

### Alternative Models

[

Salesforce/blip2-opt-2.7b

Scene Captioning

](/model/Salesforce/blip2-opt-2.7b)[

microsoft/Florence-2-large

Scene Captioning

](/model/microsoft/Florence-2-large)[

google/paligemma2-3b-mix-448

Scene Captioning

](/model/google/paligemma2-3b-mix-448)[

Qwen/Qwen3-VL-8B-Instruct

Scene Captioning

](/model/Qwen/Qwen3-VL-8B-Instruct)