TomoroAI/tomoro-colqwen3-embed-4b is a state-of-the-art ColPali-style multimodal embedding model. It maps text queries, visual documents (images, PDFs) or short videos into aligned multi-vector embeddings.

Built by merging Qwen/Qwen3-VL-4B-Instruct with Qwen/Qwen3-Embedding-4B, this model inherits robust text retrieval capabilities while preserving a full vision stack. It has been fine-tuned on a curated mixture of VDR, ViDoRe-ColPali-Training, VisRAG-Ret-Train-Synthetic-data, and VisRAG-Ret-Train-In-domain-data. It achieves SOTA or competitive performance across ViDoRe V1-V3 (English and Multilingual) while offering a significantly reduced embedding footprint compared to other full-dim Colpali model alternatives.

🛠️ Model Specifications

Feature	Detail
Architecture	Qwen3-VL 4B (Encoder-only variant) + 320-dim Projection Head
Methodology	ColPali-style Late Interaction (MaxSim scoring)
Token Budget	Up to 1,280 visual tokens per page or 5120 visual tokens per video (text prompts constrained only by the base context window)
Context Window	32k (inherited from base), typical usage < 2k tokens
Output	Multi-vector (Seq_Len × 320), L2-normalized
Supported Modalities	Text Queries, RGB Images, Synthetic Documents, Short Video (Frame-wise)
Precision	`bfloat16` weights, FlashAttention 2 enabled

Key Properties

Merged Encoders: Combines the Qwen3-VL vision encoder (patch-grid tokens with spatial merge) and language encoder.
Projection: A custom 320-dim head projects every token (text or visual) into a vector.
Processing:
- Queries: Left-padded text sequences.
- Documents: Rendered with a lightweight vision prompt and flattened into image tokens.
- Video: Supports video retrieval by decoding videos into frames and processing via the vision stack (generalization capability, not explicitly fine-tuned; dedicated benchmark coming soon).
Storage Efficiency:
- Baseline (NVIDIA Nemo-3B): Stores 1,802 tokens @ 3,072 dims (≈10.3 TB for 1M images).
- Tomoro ColQwen3: Stores max 1,280 tokens @ 320 dims (≈0.82 TB for 1M images).
- Result: 13× smaller footprint with higher performance.

📊 Evaluation Results

We report results on the ViDoRe benchmark suite. The model sets new standards on multilingual and English splits on ViDoRe V2 and V3 while maintaining comparable high performance on ViDoRe V1.

ViDoRe V3 (Latest)

English nDCG@5

Model	CompSci	Energy	FinanceEn	FinanceFr	HR	Industrial	Pharma	Physics	Avg
tomoro-colqwen3-8b	0.7443	0.6491	0.6823	0.4546	0.6421	0.5766	0.6665	0.4747	0.6113
tomoro-colqwen3-4b	0.7419	0.6023	0.6753	0.4202	0.6037	0.5787	0.6612	0.4640	0.5934
nemo-colembed-3b	0.7514	0.5838	0.6712	0.3730	0.6256	0.5447	0.6524	0.4128	0.5769
jinaai/jina-embeddings-v4	0.7175	0.5842	0.6417	0.3859	0.6206	0.5443	0.6303	0.4191	0.5680
nomic-ai/colnomic-embed-multimodal-7b	0.7528	0.5824	0.6041	0.3877	0.6060	0.5229	0.6226	0.4423	0.5651

Multilingual nDCG@5 (Excluding English Subsets)

Model	CompSci	Energy	FinanceEn	FinanceFr	HR	Industrial	Pharma	Physics	Avg
tomoro-colqwen3-8b	0.7194	0.6619	0.6172	0.4570	0.6097	0.5164	0.6403	0.4706	0.5866
tomoro-colqwen3-4b	0.7213	0.6374	0.6019	0.4305	0.5637	0.5131	0.6351	0.4636	0.5708
nemo-colembed-3b	0.7216	0.5901	0.5646	0.4102	0.5504	0.4335	0.6170	0.4192	0.5383
jinaai/jina-embeddings-v4	0.6843	0.6036	0.5482	0.4249	0.5542	0.4732	0.6059	0.4381	0.5416
nomic-ai/colnomic-embed-multimodal-7b	0.7333	0.6160	0.5219	0.4169	0.5494	0.4764	0.5938	0.4449	0.5441

ViDoRe V2

English nDCG@5

Model	BioMed	ESG HL	ESG Rpts	Economics	Avg
tomoro-colqwen3-8b	0.6784	0.7598	0.6549	0.6159	0.6772
tomoro-colqwen3-4b	0.6718	0.7465	0.6300	0.5910	0.6598
nemo-colembed-3b	0.6518	0.7538	0.6030	0.6619	0.6676
jinaai/jina-embeddings-v4	0.6359	0.6512	0.5194	0.5955	0.6005
nomic-ai/colnomic-embed-multimodal-7b	0.6479	0.6871	0.5498	0.5955	0.6201

Multilingual nDCG@5

Model	BioMed	ESG Rpts	Economics	Avg
tomoro-colqwen3-8b	0.6467	0.5911	0.5875	0.6085
tomoro-colqwen3-4b	0.6478	0.6226	0.5536	0.6080
nemo-colembed-3b	0.6187	0.5640	0.5506	0.5778
jinaai/jina-embeddings-v4	0.5994	0.5178	0.5364	0.5512
nomic-ai/colnomic-embed-multimodal-7b	0.6224	0.5336	0.5433	0.5664

ViDoRe V1 (English nDCG@5)

Model	ArxivQA	DocVQA	InfoVQA	Shift	Syn-AI	Syn-Eng	Syn-Gov	Syn-Health	TabFQuAD	Tatdqa	Avg
tomoro-colqwen3-8b	0.9115	0.6637	0.9448	0.8789	0.9926	0.9671	0.9758	0.9906	0.9423	0.8092	0.9076
tomoro-colqwen3-4b	0.9066	0.6624	0.9429	0.8739	0.9926	0.9691	0.9717	0.9963	0.9433	0.7983	0.9057
nemo-colembed-3b	0.8835	0.6621	0.9492	0.9070	0.9963	0.9663	0.9782	0.9926	0.9594	0.8057	0.9100
jinaai/jina-embeddings-v4	0.8846	0.6014	0.9379	0.9293	0.9926	0.9726	0.9659	0.9913	0.9560	0.8035	0.9035
nomic-ai/colnomic-embed-multimodal-7b	0.8832	0.6011	0.9221	0.8930	0.9876	0.9626	0.9592	0.9926	0.9596	0.8108	0.8972

Video Retrieval Evaluation

To demonstrate that Tomoro ColQwen3 strongly generalizes to video retrieval, we evaluated the models on the CareBench for text to video (General Retrieval) task and MMEB-V2 video_ret benchmark.

CareBench Evaluation

For this evaluation, we utilized a raw video encoding approach: our models encoded the video files directly without any additional textual annotations or metadata inputs. This highlights the model's ability to perform retrieval based purely on visual semantics.

Model	Recall@1	Recall@5	Recall@10
tomoro-colqwen3-8b	0.8670	0.9590	0.9850
tomoro-colqwen3-4b	0.8620	0.9570	0.9800
Care7B	0.7700	0.9560	0.9870

MMEB-V2 video_ret Evaluation

All below evaluations are using Hit@1 metric.

Model	MSR-VTT	MSVD	DiDeMo	VATEX	YouCook2	Average
tomoro-colqwen3-8b	50.3	71.2	58.8	48.0	27.8	51.2
tomoro-colqwen3-4b	51.1	72.3	59.5	49.0	26.6	51.7
IFM-TTE-7B	52.7	73.1	49.7	51.5	31.6	51.7
seed-1.6-embedding	55.3	71.3	56.7	48.8	24.6	51.3

IFM-TTE-7B and seed-1.6-embedding utilize video-text fine-tuning, whereas the Tomoro ColQwen series relies solely on image-text data.

💻 Usage

The processor exposes process_texts, process_images, and score_multi_vector.

Prerequisites

We strongly suggest flash-attn to be installed. If not, please change to attention_impl="sdpa"

Currently we only support torch==2.8.0, for higher pytorch version, please build flash attention manually, otherwise performance throughput could be low.

bash

pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install transformers pillow requests
pip install flash-attn --no-build-isolation

Inference Code

python

import torch
from transformers import AutoModel, AutoProcessor
from PIL import Image, UnidentifiedImageError
import requests
from io import BytesIO

# Configuration
MODEL_ID = "TomoroAI/tomoro-colqwen3-embed-4b"
DTYPE = torch.bfloat16
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Load Model & Processor
processor = AutoProcessor.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    max_num_visual_tokens=1280,
)
model = AutoModel.from_pretrained(
    MODEL_ID,
    dtype=DTYPE,
    attn_implementation="flash_attention_2",
    trust_remote_code=True,
    device_map=DEVICE,
).eval()

# Sample Data
queries = [
    "Retrieve the city of Singapore",
    "Retrieve the city of Beijing",
    "Retrieve the city of London",
]
docs = [
    "https://upload.wikimedia.org/wikipedia/commons/2/27/Singapore_skyline_2022.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/6/61/Beijing_skyline_at_night.JPG",
    "https://upload.wikimedia.org/wikipedia/commons/4/49/London_skyline.jpg",
]

def load_image(url: str) -> Image.Image:
    # Some CDNs (e.g., Wikimedia) expect a browser-like UA to avoid 403s.
    for headers in ({}, {"User-Agent": "Mozilla/5.0 (compatible; ColQwen3-demo/1.0)"}):
        resp = requests.get(url, headers=headers, timeout=10)
        if resp.status_code == 403:
            continue
        resp.raise_for_status()
        try:
            return Image.open(BytesIO(resp.content)).convert("RGB")
        except UnidentifiedImageError as e:
            raise RuntimeError(f"Failed to decode image from {url}") from e
    raise RuntimeError(f"Could not fetch image (HTTP 403) from {url}; try downloading locally and loading from file path.")

# Helper Functions
def encode_queries(texts, batch_size=8):
    outputs = []
    for start in range(0, len(texts), batch_size):
        batch = processor.process_texts(texts=texts[start : start + batch_size])
        batch = {k: v.to(DEVICE) for k, v in batch.items()}
        with torch.inference_mode():
            out = model(**batch)
            vecs = out.embeddings.to(torch.bfloat16).cpu()
        outputs.extend(vecs)
    return outputs

def encode_docs(urls, batch_size=4):
    pil_images = [load_image(url) for url in urls]
    outputs = []
    for start in range(0, len(pil_images), batch_size):
        batch_imgs = pil_images[start : start + batch_size]
        features = processor.process_images(images=batch_imgs)
        features = {k: v.to(DEVICE) if isinstance(v, torch.Tensor) else v for k, v in features.items()}
        with torch.inference_mode():
            out = model(**features)
            vecs = out.embeddings.to(torch.bfloat16).cpu()
        outputs.extend(vecs)
    return outputs

# Execution
query_embeddings = encode_queries(queries)
doc_embeddings = encode_docs(docs)

# MaxSim Scoring
scores = processor.score_multi_vector(query_embeddings, doc_embeddings)
print(scores)

🎞️ Lightweight Video Retrieval

ColQwen3 generalizes to short videos while learning from image-text retrieval task. This minimal example samples a clip with torchvision, encodes queries and frames, then pools frame embeddings with a per-dimension max before MaxSim scoring.

We recommand use of maximum 5120 visual tokens for video retrieval task for best performance.

python

from pathlib import Path

import torch
from transformers import AutoModel, AutoProcessor

MODEL_ID = "TomoroAI/tomoro-colqwen3-embed-4b"
DTYPE = torch.bfloat16
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

processor = AutoProcessor.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    max_num_visual_tokens=5120,
)
model = AutoModel.from_pretrained(
    MODEL_ID,
    dtype=DTYPE,
    attn_implementation="flash_attention_2",
    trust_remote_code=True,
    device_map=DEVICE,
).eval()

queries = ["Retrieve the football video", "Find the basketball clip", "Find the swimming clip", "Find the wrestling clip"]
videos = ["/root/sample_videos/football.mp4", "/root/sample_videos/basketball.mp4", "/root/sample_videos/swimming.mp4",  "/root/sample_videos/wrestling.mp4"]


def encode_queries(texts):
    batch = processor.process_texts(texts=texts)
    batch = {k: v.to(DEVICE) for k, v in batch.items()}
    with torch.inference_mode():
        out = model(**batch)
    return out.embeddings.to(torch.bfloat16).cpu()

def encode_videos(paths):
    vids = [str(Path(p).expanduser()) for p in paths]
    feats = processor(
        videos=vids,
        padding="longest",
        return_tensors=None,  # keep metadata as Python objects until we drop it
        videos_kwargs={"return_metadata": True},
    )
    feats.pop("video_metadata", None)  # drop metadata before forwarding to the model
    feats = feats.convert_to_tensors(tensor_type="pt")
    feats = {k: v.to(DEVICE) if isinstance(v, torch.Tensor) else v for k, v in feats.items()}
    with torch.inference_mode():
        out = model(**feats)
    return out.embeddings.to(torch.bfloat16).cpu()

q_emb = encode_queries(queries)
v_emb = encode_videos(videos)
scores = processor.score_multi_vector(q_emb, v_emb)
print(scores)

⚖️ Strengths & Limitations

Strengths

Performance: State of the art retrieval performance on ViDoRe V2 & V3 dataset with excellent performance on multimodal document retrieval.
Complex Layouts: Excellent handling of chart-rich PDFs, domain-specific documents.
End-to-end Retrieval: Capable of OCR-free retrieval on unseen multimodal documents without using an intermediate vision LLM to generate summary for retrieval.
Retrieval Task Transfer: Inherited strong text retrieval performance from the merged vector of the Qwen3-Embedding-8B model.
Multilingualism: Strong performance on non-English document inputs.

Limitations

Video Support: The retrieval model generalizes to video retrieval on our preliminary findings, however it's not fine-tuned on large-scale video retrieval datasets, we plan to further improve this in the future.
Storage Cost: Still larger than single‑vector baselines despite the smaller token dimension.
Retrieval Instructions: The model currently is not fine-tuned with diverse system instructions similar to Qwen3-Embedding models, we intent to improve this with more synthetic dataset in the future.

License & Data

Distributed under Apache 2.0.

Weights: Upstream Qwen checkpoints retain their community licenses; ensure compliance when mixing.
Data: Training data includes ViDoRe/MTEB corpora and synthetic VisRAG assets.

Acknowledgement

We gratefully acknowledge the support of Tomoro AI, a leading AI engineering firm dedicated to delivering high-quality enterprise solutions that accelerate complex R&D and business transformation. This work is directly applied to enhance Tomoro’s customized multimodal agentic RAG pipelines, empowering the autonomous agents to parse, reason over, and retrieve from large-scale enterprise internal documentation. By bridging the gap between vision and language, this model supports Tomoro AI's mission to accelerate the delivery of high-quality enterprise multimodal solutions and deploy robust, production-grade intelligence across high-stakes industries.

📚 Citation

If you use this model, please cite:

bibtex

@misc{huang2025beyond,
  author = {Huang, Xin and Tan, Kye Min},
  title = {Beyond Text: Unlocking True Multimodal, End-to-end RAG with Tomoro ColQwen3},
  year = {2025},
  url = {https://tomoro.ai/insights/beyond-text-unlocking-true-multimodal-end-to-end-rag-with-tomoro-colqwen3},
  publisher = {Tomoro.ai}
}

Johnny667/tomoro-colqwen3-embed-4b

TomoroAI/tomoro-colqwen3-embed-4b

⚡ Executive Summary