ShayonSarker/xglm-2.9B-GGUF

<div align="center">

<p> <a href="https://huggingface.co/ShayonSarker/xglm-2.9B-GGUF"><img alt="Hugging Face GGUF" src="https://img.shields.io/badge/Hugging%20Face-GGUF-FFD21E?style=for-the-badge"></a> <img alt="XGLM model" src="https://img.shields.io/badge/Model-XGLM--2.9B-8A2BE2?style=for-the-badge"> <img alt="GGUF formats" src="https://img.shields.io/badge/GGUF-F16%20%7C%20Q80%20%7C%20Q4K_M-FFD21E?style=for-the-badge"> <img alt="Languages" src="https://img.shields.io/badge/Languages-30%2B-00A6A6?style=for-the-badge"> <img alt="Bengali and Bangla" src="https://img.shields.io/badge/Bengali-Bangla-16A34A?style=for-the-badge"> <img alt="Native llama.cpp" src="https://img.shields.io/badge/llama.cpp-native-24292F?style=for-the-badge"> <img alt="Validated" src="https://img.shields.io/badge/validated-pass-22C55E?style=for-the-badge"> <img alt="MIT license" src="https://img.shields.io/badge/License-MIT-7C3AED?style=for-the-badge"> </p>

🌍 XGLM-2.9B GGUF

Meta’s multilingual XGLM base model for higher-quality Bengali and global language workloads.

🌐 30+ Languages &nbsp;•&nbsp; 🇧🇩 Bengali / Bangla &nbsp;•&nbsp; 🧠 Base Model &nbsp;•&nbsp; ⚙️ Native GGUF &nbsp;•&nbsp; 📦 2.9B Parameters &nbsp;•&nbsp; ⚖️ MIT

Source model · llama.cpp

👇 View verified English and Bangla question/answer examples

</div>

✨ Highlights

• —Native xglm architecture support for llama.cpp
• —Faithful XGLM embeddings, positional encoding, attention, GELU, and tokenizer behavior
• —256,008-token vocabulary with verified multilingual token parity
• —2,048-token context window
• —F16, Q80, and importance-matrix-calibrated Q4K_M formats

📦 Choose a Format

🏗️ Rebuild

The GitHub release hub includes the end-to-end builder, native patch, verifier, and pinned dependencies.

git clone https://github.com/Dadhichi-Sarker-Shayon/XGLM-2.9B-GGUF.git
cd XGLM-2.9B-GGUF
python -m pip install -r requirements-build.txt
python build_gguf.py

The 2.9B build requires at least 30 GB of free disk space. It does not overwrite this release.

🚀 Run Locally

The included `xglm-llama.cpp.patch` adds native XGLM support to llama.cpp.

hf download ShayonSarker/xglm-2.9B-GGUF xglm-llama.cpp.patch --local-dir .
hf download ShayonSarker/xglm-2.9B-GGUF XGLM-2.9B-Q4_K_M.gguf --local-dir .

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git checkout 6b790a9c291b5d7af3312bbf9f0c558aa023b13e
git apply ../xglm-llama.cpp.patch
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --target llama-completion

./build/bin/llama-completion \
  -m ../XGLM-2.9B-Q4_K_M.gguf \
  -p "The capital of France is" \
  -n 32 --temp 0

<a id="verified-question-answer-examples"></a>

❓ Verified Question → Answer Examples

These are actual Q4KM completions at temperature 0. The answers below were fact-checked; incorrect, repetitive, and evasive completions were removed. This curated set is not a benchmark. XGLM is a base model, so the questions are intentionally simple.

📈 Performance

Lower perplexity (PPL) is better. Scores use separate held-out English and Bengali text with 64-token evaluation windows.

🔬 Validation

• —Token IDs match Transformers across Bengali, English, French, Chinese, and Arabic.
• —F16 English and Bengali predictions match Transformers.
• —Both quantized formats pass the English and Bengali quality gates.

🧩 Intended Use

XGLM-2.9B is a base language model, not an instruction-tuned assistant. It is suitable for higher-quality multilingual research, Bengali/English workloads, local generation, and GGUF runtime testing.

Outputs may be inaccurate or inappropriate. Validate important results independently.

📄 License

MIT. See the upstream model card for source-model details and attribution.