Skip to content
Other

ColQwen3.5-4.5B-v4

visual document retrieval powered by Qwen3.5-4B

2 min readupdated

ColQwen3.5-4.5B-v4 by athrael-soju tackles visual document retrieval by matching text queries directly against PDF pages, charts, and scans without running OCR first. Built on the Qwen3.5-4B architecture with bidirectional attention, the 4.54B-parameter model uses late interaction: both queries and document pages are encoded into sets of token vectors, and relevance is scored via MaxSim. Unlike the base model, it is specialized for multi-vector search and supports Matryoshka embeddings (a technique that lets you truncate vectors without a major drop in accuracy). It is well suited for developers building multimodal RAG pipelines and enterprise search engines over complex reports.

Key features

  • Flexible vector dimensions: the projection layer is trained with Matryoshka prefixes at 64, 128, 256, 512, 1024, and 2048 dimensions, cutting index size by 2x to 32x.
  • Multi-stage distillation: v4 is trained using a larger teacher model and a reranker that filters out false negatives, combined via LoRA adapter merging.
  • Six-language support: model metadata lists support for English, French, German, Spanish, Italian, and Portuguese.
  • Permissive license: released under the Apache 2.0 license and compatible with the sentence-transformers library.

What you can do

  • Plug it into a multimodal RAG pipeline to retrieve answers from financial, medical, and scientific reports packed with complex tables and charts.
  • Build a compact vector index for massive scan archives by truncating embeddings down to 64 or 256 dimensions to save memory.
  • Rank document pages against text queries without an intermediate OCR parsing step.

Example

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("athrael-soju/ColQwen3.5-4.5B-v4")
queries = ["Total outlay is maximum in which year?"]
images = ["[https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg](https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg)"]

q_emb = model.encode_query(queries)
i_emb = model.encode_document(images)
scores = model.similarity(q_emb, i_emb)
Read moreDetails
Benchmarks (nDCG, 2048-d, 1792 visual tokens): ViDoRe V1 Mean — 93.09 (nDCG@5), ViDoRe V2 Mean — 69.78 (nDCG@5), ViDoRe V3 Mean — 67.14 (nDCG@10).
Software requirements: requires transformers>=5.2.0 and sentence-transformers[image]>=6.0.0.
Limitations: older versions of transformers ignore the is_causal: false flag and produce corrupted embeddings.
Weights format: BF16, 4.5B parameters.
Discuss Telegram 2 30

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.5 GB · ready files9.8 GB
  • FP8 / INT8recommendedweights 4.2 GB5.2 GB
  • 4-bitweights 2.4 GB3.1 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 2060 · BF16 — from RTX 3080

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

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.

    # athrael-soju/ColQwen3.5-4.5B-v4
    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 athrael-soju/ColQwen3.5-4.5B-v4 --local-dir ~/models/ColQwen3.5-4.5B-v4
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to ColQwen3.5-4.5B-v4
  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

  1. Qwen3.5-4B-Basefamily root
  2. Qwen3.5-4B
  3. ColQwen3.5-4.5B-v4

New posts about versions of Qwen3.5-4B-Base (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

athrael-soju/ColQwen3.5-4.5B-v4
Author
athrael-soju
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

302

Ask AI about this model

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

Comments