bidirlm-omni-2.5b-textonly-GGUF

bidirlm-omni-2.5b-textonly GGUF

GGUF format of BidirLM/BidirLM-Omni-2.5B-Embedding for use with CrispEmbed.

BidirLM-Omni 2.5B — text-only GGUF. Qwen3-derived bidirectional encoder, 2048-d shared embedding space, 90+ languages. The upstream model's audio + vision towers are NOT included in this file (use the bidirlm-omni-2.5b-GGUF variant for cross-modal embedding).

Note: text-only GGUF

The upstream model is omnimodal (text + image + audio). This GGUF contains only the text path — the bidirectional Qwen3 body with mean pooling, producing 2048-d embeddings in the model's shared cross-modal space.

For text → text similarity (semantic search, retrieval, clustering across 90+ languages), this GGUF is functionally complete and matches the upstream reference at cosine ≥ 0.999.

For cross-modal queries (text ↔ image, text ↔ audio), use the bidirlm-omni-2.5b-GGUF variant — it includes the audio tower (text + audio cross-modal) and is the recommended choice for omnimodal retrieval. The vision tower is not yet supported by either GGUF.

Files

Parity vs HuggingFace reference

Cosine similarity vs the upstream sentence-transformers reference on a fixed test set (text, jfk.wav for audio):

Note: below the 0.99 retrieval-quality bar — text: q5_k (0.983), q4_k (0.937). Embeddings are still functionally usable (>0.9 = directionally correct for similarity ranking) but expect small differences in nearest-neighbor results vs the upstream f32 reference.

Quick Start

# Download
huggingface-cli download cstr/bidirlm-omni-2.5b-textonly-GGUF bidirlm-omni-2.5b-textonly-f16.gguf --local-dir .

# Run with CrispEmbed
./crispembed -m bidirlm-omni-2.5b-textonly-f16.gguf "Hello world"

# Or with auto-download
./crispembed -m bidirlm-omni-2.5b-textonly "Hello world"

Model Details

Verification

Verified bit-identical to HuggingFace sentence-transformers (cosine similarity >= 0.999 on test texts).

Usage with CrispEmbed

CrispEmbed is a lightweight C/C++ text embedding inference engine using ggml. No Python runtime, no ONNX. Supports BERT, XLM-R, Qwen3, and Gemma3 architectures.

# Build CrispEmbed
git clone https://github.com/CrispStrobe/CrispEmbed
cd CrispEmbed
cmake -S . -B build && cmake --build build -j

# Encode
./build/crispembed -m bidirlm-omni-2.5b-textonly-f16.gguf "query text"

# Server mode
./build/crispembed-server -m bidirlm-omni-2.5b-textonly-f16.gguf --port 8080
curl -X POST http://localhost:8080/v1/embeddings \
    -d '{"input": ["Hello world"], "model": "bidirlm-omni-2.5b-textonly"}'

Credits

• Original model: BidirLM/BidirLM-Omni-2.5B-Embedding
• Inference engine: CrispEmbed (ggml-based)
• Conversion: convert-decoder-embed-to-gguf.py

Provenance and EU AI Act Art. 53 note

• Upstream model: BidirLM/BidirLM-Omni-2.5B-Embedding — published by BidirLM.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.