---

license: mit

base_model: facebook/xglm-2.9B

tags:

- gguf

- xglm

- multilingual

- bengali

- bangla

- text-generation

language:

- bn

- en

- fr

- de

- ar

- ru

- zh

- ja

- es

---

<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%20Q8_0%20%7C%20Q4_K_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="Runtime" src="https://img.shields.io/badge/runtime-llama.cpp%20%2B%20patch-E8590C?style=for-the-badge">

<img alt="Validated" src="https://img.shields.io/badge/validated-pass-22C55E?style=for-the-badge">

<img alt="F16 parity with transformers" src="https://img.shields.io/badge/F16%20parity-12%2F14%20exact-00A6A6?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; ⚙️ GGUF (llama.cpp + patch) &nbsp;•&nbsp; 📦 2.9B Parameters &nbsp;•&nbsp; ⚖️ MIT

Source model · llama.cpp

👇 View verified English and Bangla question/answer examples

</div>

---

✨ Highlights

• XGLM support for llama.cpp via the included xglm-llama.cpp.patch (upstream does not ship this architecture)
• 256,008-token vocabulary with verified multilingual token parity
• 2,048-token context window
• F16, Q8_0, and importance-matrix-calibrated Q4_K_M formats

📦 Choose a Format

| File | Status | Best for |

|---|---|---|

| XGLM-2.9B-F16.gguf | Published | Reference quality and maximum fidelity |

| XGLM-2.9B-Q8_0.gguf | Published | Strong quality with lower memory use |

| XGLM-2.9B-Q4_K_M.gguf | Published | Strongest compact option for local inference |

⚠️ Runtime requirement

Stock llama.cpp cannot load these files and fails with unknown model architecture: 'xglm'. XGLM is not in upstream llama.cpp. Apply the bundled patch against the pinned commit:

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

Alternatives that need no patch: transformers with facebook/xglm-2.9B, or the HF gguf file with a runtime that implements xglm.

🏗️ Rebuild

The GitHub release hub includes the end-to-end builder, XGLM runtime 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

Apply the runtime patch first, then:

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

llama-completion -m ./XGLM-2.9B-Q4_K_M.gguf \
  -p "Question: What is the capital of Japan?
Answer:" \
  -n 24 --temp 0

Expected output: Tokyo.

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

❓ Verified Question → Answer Examples

Every row below is a verbatim XGLM-2.9B-Q4_K_M.gguf completion from the patched build at --temp 0, 24 new tokens. Nothing is hand-written; wrong output is published in the failure table below rather than edited into looking right. This is a smoke test, not a benchmark.

Prompt form Question: ...\nAnswer:

| Question | Model answer (verbatim) |

|---|---|

| What is the capital of Japan? | Tokyo. |

| What is the capital of Italy? | Rome. |

| What is the capital of Egypt? | Cairo. |

| What is the largest ocean on Earth? | The Pacific Ocean. |

| Which planet is closest to the Sun? | Mercury. |

| How many continents are there? | There are seven continents. |

| What is the chemical symbol for gold? | Gold is a chemical element with symbol Au (from Latin: aurum). |

Plain completion form (a base model, so sentence continuation often works better than Q/A)

| Prompt | Model answer (verbatim) |

|---|---|

| The capital city of France is | Paris. |

| The chemical symbol for water is | H2O. |

| The capital city of Bangladesh is | also known as the capital of the country. Dhaka is the capital of Bangladesh. |

| Plants absorb | carbon dioxide from the atmosphere and convert it into sugars, which are then used by the plant to produce energy. |

🔍 Is a wrong answer our bug or the model's? We tested it

14 prompts (8 English Q&A, 3 plain completions, 3 Bengali) were run through facebook/xglm-2.9b in transformers and through our GGUFs — all greedy, 24 new tokens, same prompt strings.

| Format | Exact string match vs transformers |

|---|---:|

| F16 | 12 / 14 |

| Q4_K_M | 8 / 14 |

Both F16 differences are 24-token truncation points rather than errors: the two outputs start identically and diverge mid-sentence.

The conversion is faithful. F16 reproduces the reference model string-for-string, including the answers that are factually wrong. A broken conversion or a broken runtime patch could not track the reference this closely. Any XGLM factual error you see below is the upstream checkpoint's own behaviour.

⚠️ Why Q4_K_M looks worse than F16

| Prompt | transformers reference | F16 (ours) | Q4_K_M (ours) |

|---|---|---|---|

| The largest planet in the Solar System is | Jupiter. It is the largest planet in the Solar System... | Jupiter. It is the largest planet in the Solar System... | the Earth. |

| Question: How many days are in a leap year?\nAnswer: | The leap year is a year that is longer than the normal year. | The leap year is a year that begins on a leap day... | There are 30 days in a leap year. |

| প্রশ্ন: জাপানের রাজধানী কোন শহর?\nউত্তর: | জাপানের রাজধানী হচ্ছে – টোকিও। | জাপানের রাজধানী হচ্ছে – টোকিও। | loops the prompt and loses the answer |

| প্রশ্ন: সোনার রাসায়নিক প্রতীক কী?\nউত্তর: | সোনার রাসায়নিক প্রতীক হলো সোনালি রঙের লাল চিহ্ন। | সোনার রাসায়নিক প্রতীক হলো সোনালি রঙের লাল চিহ্ন। | সোনার রাসায়নিক প্রতীক হলো Sn। |

Use F16 or Q8_0 for Bengali and for factual lookups. Q4_K_M is fine for bulk text generation but it is the first thing to break on lower-resource languages.

Perplexity does not predict this. The table below shows Q4_K_M with better English PPL than F16 (82.88 vs 91.22) while being measurably less accurate on the prompts above. Treat PPL as a text-fluency proxy, not an accuracy metric.

⚠️ Known Failures (Q4_K_M, the published default)

These are the exact completions the released Q4_K_M file produces:

| Prompt | Model answer (verbatim) |

|---|---|

| Question: How many days are in a leap year?\nAnswer: | There are 30 days in a leap year. |

| The largest planet in the Solar System is | the Earth. |

| There are | no reviews yet. |

| Question: What is the capital of Japan?\nAnswer: (Bengali script) | জাপানের রাজধানী হচ্ছে জাপানের রাজধানী হচ্ছে জাপানের রাজধানী হচ্ছে |

| সূর্যের নিকটতম গ্রহ কোনটি? | সূর্যের নিকটতম গ্রহ হচ্ছে- মঙ্গল। |

| ফ্রান্সের রাজধানী | প্যারিসে ঐতিহাসিক স্থাপনাগুলোর মধ্যে অন্যতম হলো প্যারিস প্যারেড। |

Bengali/Bangla prompting is unreliable in this model family, and Q4_K_M makes it worse. The reference model also gets several of these wrong, so treat Bangla output as a research baseline, not a finished translation or QA system.

📈 Performance

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

| Format | English PPL | vs F16 | Bengali PPL | vs F16 |

|---|---:|---:|---:|---:|

| F16 | 91.22 | Baseline | 8.06 | Baseline |

| Q8_0 | 87.35 | -4.24% | 8.00 | -0.78% |

| Q4_K_M | 82.88 | -9.15% | 8.52 | +5.70% |

🔬 Validation

• Generation parity with transformers: 14 greedy prompts (8 English Q&A, 3 plain completions, 3 Bengali) — F16 matches the reference output exactly on 12/14, with the 2 differences being 24-token truncation points rather than content errors. Q4_K_M matches on 8/14. See the parity section above.
• Token IDs match Transformers across Bengali, English, French, Chinese, and Arabic.
• Both quantized formats pass the English and Bengali held-out perplexity gates.
• Not tested: logit-level numeric parity, and long-context behaviour beyond 24 generated tokens.

🧩 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.