anushkrishna/falcon-models

Falcon model packs

Prepared English/Arabic model artifacts for the Falcon desktop application. This is a multi-pack distribution repository, not a single Transformers model. No Hugging Face account is required to download these public files.

Pack	Files	Runtime	Changes from upstream	License
ALLaM-7B-Instruct-preview	allam/	llama.cpp	Converted to GGUF and quantized Q4_K_M	Apache-2.0
OPUS-MT English → Arabic	marian/en-ar/	CTranslate2	Converted to CTranslate2 and quantized int8	CC-BY-4.0
OPUS-MT Arabic → English	marian/ar-en/	CTranslate2	Converted to CTranslate2 and quantized int8	CC-BY-4.0

ALLaM is attributed to its original authors at NCAI/SDAIA; its upstream repository is maintained under HUMAIN. OPUS-MT is attributed to Helsinki-NLP / the OPUS-MT project, University of Helsinki. See the original model cards in sources/ for authorship, citations, intended uses, and limitations. Falcon's packaging does not imply endorsement by the original authors.

manifest.json records the exact upstream revisions, artifact sizes and SHA-256 hashes. Original model cards are preserved in sources/; applicable license texts and attribution are linked in LICENSE.md. The transformations are format conversion and quantization, with no additional fine-tuning. The exact Marian converter environment was not recorded for these existing artifacts; their identity is fixed by hashes. Future conversions should record the complete toolchain versions before publishing a new release.

Downloading

Use a full commit SHA from the release history, rather than main or a mutable tag, and verify the file size and SHA-256 before loading. For example:

https://huggingface.co/anushkrishna/falcon-models/resolve/<commit-sha>/marian/en-ar/model.bin
https://huggingface.co/anushkrishna/falcon-models/resolve/<commit-sha>/allam/allam-7b-instruct-preview-Q4_K_M.gguf

Marian requires all five files in the selected direction folder. ALLaM's tokenizer/config files are supplied alongside its GGUF for provenance and compatibility. The GGUF includes the tokenizer and chat template.

OCR continues to use the official PaddlePaddle Hugging Face repositories; external_artifacts in the manifest documents Falcon's pinned OCR weights. Those weights and native runtime binaries are not mirrored in this release.

Falcon downloads model data during setup and runs document processing locally. Model updates should be tested and explicitly approved by the application release; uploading new files here does not update an already pinned app.