Skip to content
Other

PEARL-Qwen3-VL-8B-LoRA

an adapter for streaming video analysis using custom concepts

2 min readupdated

This is a LoRA (a lightweight add-on for a base model) for the Qwen3-VL-8B-Instruct vision-language model, trained as part of the PEARL framework. It is built for personalized streaming video understanding, letting users ask questions about custom concepts defined at runtime. The model handles both questions about the current video segment (Real-Time) and memory-based queries from earlier segments (Past-Time). Applied during the final response stage alongside a retriever, the adapter boosts overall accuracy.

Key features

  • Trained on the PSVU-10K dataset, which contains 10,000 questions across videos from 155 sources.
  • Improves response accuracy on PSVU-Bench across all categories, including behavior, objects, and locations.
  • Contains 15.3M trainable parameters with rank 16 and alpha 32.
  • Released under the permissive Apache-2.0 license.

What you can do

  • Use it in the PEARL pipeline to answer questions about custom objects in streaming video.
  • Load the adapter weights in vLLM with LoRA support enabled to generate final responses.
  • Integrate the adapter via transformers and peft for custom video processing pipelines.

Example

from vllm import LLM, SamplingParams
from huggingface_hub import snapshot_download
from vllm.lora.request import LoRARequest

llm = LLM(model="Qwen/Qwen3-VL-8B-Instruct", enable_lora=True, max_lora_rank=16)
answer_lora = LoRARequest("pearl-answer", 1, snapshot_download("zyh200727/PEARL-Qwen3-VL-8B-LoRA"))
sampling = SamplingParams(temperature=0.0)

helper_outputs = llm.generate(helper_inputs, sampling)
answer_outputs = llm.generate(answer_inputs, sampling, lora_request=answer_lora)
Read moreDetails
Base model: Qwen/Qwen3-VL-8B-Instruct
Overall accuracy on PSVU-Bench: 61.41% (with adapter)
Training: 9,140 steps, AdamW, batch size 8, BF16
Weights: FP32, 61.4 MB
Discuss Telegram 0 24

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 16 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.2 GB9.5 GB
  • 4-bitweights 4.6 GB5.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 3080 · BF16 — from RTX 4090

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model Qwen3-VL-8B-Instruct.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # zyh200727/PEARL-Qwen3-VL-8B-LoRA
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Qwen/Qwen3-VL-8B-Instruct --local-dir ~/models/Qwen3-VL-8B-Instruct
    hf download zyh200727/PEARL-Qwen3-VL-8B-LoRA --local-dir ~/models/PEARL-Qwen3-VL-8B-LoRA
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3-VL-8B-Instruct
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

Qwen/Qwen3-VL-8B-Instruct

This model is a LoRA / adapter. Download the base model first, then load this one on top.

New posts about versions of Qwen3-VL-8B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face
Merges 1All on Hugging Face

About the model

zyh200727/PEARL-Qwen3-VL-8B-LoRA
Author
zyh200727
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments