BidirLM/BidirLM-Omni-2.5B-Embedding

BidirLM-Omni-2.5B

BidirLM-Omni is the omnimodal variant of the BidirLM family — a 2.5B bidirectional encoder that jointly embeds text, images, and audio into a shared representation space, enabling state-of-the-art embedding performance.

Supported Tasks

Multimodal embeddings (via Sentence Transformers): cross-modal retrieval (text ↔ image, text ↔ audio), multimodal semantic similarity, clustering, and classification across text, image, and audio modalities.

Text-only downstream fine-tuning (via Transformers): sequence classification (e.g. MNLI, XNLI), token classification (e.g. NER), sequence regression.

Supported Languages Multilingual support across over 119 languages, inherited from the Qwen3 base model and reinforced through contrastive training with 87 languages.

Usage

Sentence Transformers

Pass inputs directly to encode(). All modalities produce embeddings in the same 2048-dimensional space and can be compared cross-modally.

Modality	Input type	Notes
Text	str	Any language; no length limit (model context is 32k tokens)
Image	PIL.Image.Image	Any size and aspect ratio; resized internally
Audio	np.ndarray, list[float], or dict with "array" (np.ndarray) and "sampling_rate" (int)	Any sample rate; resampled to 16 kHz internally via librosa
Mixed	list[dict] conversation (role/content)	Interleave text + image or text + audio in a single prompt — see Chat Template below

import numpy as np
import PIL.Image
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("BidirLM/omni-st5.4", trust_remote_code=True)

# Text queries
texts = [
    "An image with a red background.",
    "An image with a blue background.",
    "A deep bass sound.",
    "A high-pitched sound.",
]

# Images, synthetic solid-color 256x256 images
images = [
    PIL.Image.fromarray(np.full((256, 256, 3), (220, 30, 30), dtype=np.uint8)),  # red
    PIL.Image.fromarray(np.full((256, 256, 3), (30, 30, 220), dtype=np.uint8)),  # blue
]

# Audio, synthetic sine waves at 16kHz, 2 seconds each
sr = 16000
t  = np.linspace(0, 2.0, sr * 2, endpoint=False, dtype=np.float32)
audios = [
    {"array": np.sin(2 * np.pi *   80 * t), "sampling_rate": sr},  #   80 Hz — bass
    {"array": np.sin(2 * np.pi * 7500 * t), "sampling_rate": sr},  # 7500 Hz — high
]

# Encode all modalities and compute similarities
text_embeddings  = model.encode(texts)
image_embeddings = model.encode(images)
audio_embeddings = model.encode(audios)

# Pass a custom instruction via prompt= (applies to all items in the batch)
# text_embeddings  = model.encode(texts, prompt="Retrieve semantically similar text.")

print(model.similarity(text_embeddings, image_embeddings))
print(model.similarity(text_embeddings, audio_embeddings))

# Text-Image similarity             red img   blue img
# "An image with a red background." [ 0.6928,   0.3103]  ← high red match
# "An image with a blue background."[ 0.4278,   0.6436]  ← high blue match
# "A deep bass sound."              [ 0.1519,   0.2272]  ← low (text/image mismatch)
# "A high-pitched sound."           [ 0.1418,   0.1812]  ← low (text/image mismatch)

# Text-Audio similarity             80Hz bass  7500Hz high
# "An image with a red background." [ 0.0010,   0.0410]  ← low (image/audio mismatch)
# "An image with a blue background."[ 0.0526,   0.0642]  ← low (image/audio mismatch)
# "A deep bass sound."              [ 0.5456,   0.4243]  ← higher bass match
# "A high-pitched sound."           [ 0.4004,   0.5177]  ← higher high-pitch match

Transformers - Fine-tuning for Downstream Tasks

import numpy as np
import PIL.Image
from transformers import AutoProcessor, AutoModelForSequenceClassification, AutoModelForTokenClassification

processor = AutoProcessor.from_pretrained(
    "BidirLM/BidirLM-Omni-2.5B-Embedding", trust_remote_code=True
)

sr = 16000
conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": PIL.Image.fromarray(np.zeros((256, 256, 3), dtype=np.uint8))},
            {"type": "audio", "audio": {"array": np.zeros(sr, dtype=np.float32), "sampling_rate": sr}},
            {"type": "text",  "text": "Your text."},
        ],
    }
]
processor.apply_chat_template(conversation, tokenize=True, add_generation_prompt=False)


# Sequence classification (e.g., NLI)
seq_model = AutoModelForSequenceClassification.from_pretrained(
    "BidirLM/BidirLM-Omni-2.5B-Embedding",
    trust_remote_code=True,
    num_labels=3,
)

# Token classification (e.g., NER)
tok_model = AutoModelForTokenClassification.from_pretrained(
    "BidirLM/BidirLM-Omni-2.5B-Embedding",
    trust_remote_code=True,
    num_labels=7,
)

Requirements

transformers>=5.5.0
sentence-transformers>=5.4.0
librosa>=0.10.0

FAQ

1. What pooling strategy does this model use?

The model uses mean pooling across all modalities. This is handled automatically when using Sentence Transformers.

2. Do I need trust_remote_code=True?

Yes. BidirLM-Omni uses a custom bidirectional omnimodal architecture that requires loading custom code from the repository.

3. Can I compare embeddings across modalities?

Yes. Text, image, and audio embeddings live in the same 2048-dimensional space and can be compared directly using cosine similarity.

4. What audio formats and sample rates are supported?

Any sample rate is accepted — the model resamples internally using librosa when the source rate differs from the native 16 kHz. Three input formats are supported:

• np.ndarray — a 1-D float32 array of raw samples
• list[float] — a plain Python list of samples
• dict with "array" (np.ndarray) and "sampling_rate" (int) — the format returned by HuggingFace datasets Audio features

Any audio format readable by standard libraries (WAV, MP3, FLAC, etc.) can be used by loading it into a NumPy array first (e.g. with librosa.load or soundfile.read).

Citation

@misc{boizard2026bidirlmtextomnimodalbidirectional,
      title={BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs}, 
      author={Nicolas Boizard and Théo Deschamps-Berger and Hippolyte Gisserot-Boukhlef and Céline Hudelot and Pierre Colombo},
      year={2026},
      eprint={2604.02045},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.02045}, 
}