Skip to content
Other

CGVM-R2.6-Qwen3VL-8B-step600

checkpoint for accurate analysis of short and long videos

2 min readupdated

An intermediate LoRA adapter (a lightweight add-on on top of the base model) checkpoint for the multimodal Qwen3-VL-8B-Instruct model, created by XiangnanW. It handles video understanding and multiple-choice Video QA tasks, fixing weaknesses from the previous R2.5 release when dealing with short clips, motion direction, speed, and repetition counting. Built for researchers and video analytics developers who need consistent accuracy across both short clips under 24 seconds and long video streams.

Key features

  • Dense sampling for short videos: For clips shorter than 24 seconds, the model uniformly extracts a full 16 frames across the entire clip instead of sparse 1 FPS (frames per second) sampling, while keeping the existing handling logic for streams longer than 24 seconds.
  • Support for 2–5 multiple-choice questions: The inference pipeline uses log_softmax calibration that relies on the LoRA prediction when confidence is high and falls back to the base model only in ambiguous cases.
  • Decoding failure protection: Adds automatic fallback to sequential reading via cv2 for videos with invalid negative timestamps (PTS), preventing hangs during processing.
  • Expanded 28,000-sample dataset: Adds 8,000 question-answer pairs from PerceptionTest, MotionBenchCaption, and SSv2 to the base data, focusing on action order, camera angle, speed, and reverse playback.

What you can do

  • Run video understanding benchmarks: Test the step-600 weights alongside the updated repository code on TempCompass and MotionBench.
  • Integrate into a Video QA pipeline: Use the updated frame buffering mechanism from src/ to analyze short video clips and streaming video.
  • Compare against R2.5: Measure accuracy gains on motion direction, speed estimation, and counting repeated actions.

Example

torchrun --nproc_per_node=8 scripts/evaluate_adapted.py \
  --variant R2 --benchmark tempcompass \
  --reader-dir checkpoints/R2/reader \
  --output-root results/R2_6_step600 --resume
Read moreDetails
Weight status: intermediate snapshot step-000600 out of 875 steps (trained on 19,200 of 28,000 examples, 68.57%), 677 MB adapter size.
Training parameters: 4×H100, global_batch=32, lr=6e-5, LoRA r=64, alpha=64, kl_weight=0.10.
Preliminary benchmarks (R2.5 + new R2.6 pipeline): TempCompass (n=160) rose from 73.75% to 75.62%, MotionBench (n=156) from 57.05% to 61.54%.
Limitation: All fixes require the updated inference code from the src/ folder or the cgvm-r26-src-v1.tar.gz archive; legacy R2.5 code will not activate the new mechanisms.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 16 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.3 GB9.6 GB
  • 4-bitweights 4.7 GB5.7 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.

    # XiangnanW/CGVM-R2.6-Qwen3VL-8B-step600
    # ===== 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 XiangnanW/CGVM-R2.6-Qwen3VL-8B-step600 --local-dir ~/models/CGVM-R2.6-Qwen3VL-8B-step600
    # 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

XiangnanW/CGVM-R2.6-Qwen3VL-8B-step600
Author
XiangnanW
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments