Granite5 diagnostic conversion checkpoints โ€” NOT accepted model support

Diagnostic inputs for porting IBM Granite Speech5.0-2b to CrispASR. Original source: ibm-granite/granite-speech-5.0-2b at c075cf19e041a2e4d16c5ee24763e5b037d8f479, declared Apache-2.0 by IBM.

The pre-YaRN baselines intentionally omit required RoPE scaling. Native stage parity, decoded-output acceptance and actual CUDA are pending. A successful conversion/upload is not acceptance. No production registry points here.

Modifications: CrispASR GGUF conversion and quantization. Recipes retain their separate MIT scope. Original publisher model card, source checksums, recipes and per-run receipts are retained. Full Apache LICENSE and IBM attribution are provided. No IBM endorsement is implied.

License scopes for validation artifacts

The Apache-2.0 model-card tag describes the IBM-derived model weights. It does not relicense the validation tools or their dependencies. CrispASR code and conversion recipes retain their MIT terms. Compiled validation capsules include third-party runtime libraries under their respective licenses; preserve the capsule LICENSE, THIRD_PARTY_NOTICES.txt, per-package copyright/license notices, and accompanying corresponding-source packets. The capsule provenance identifies the exact source commits, binary packages, source versions and payload hashes.

NVIDIA driver, CUDA runtime and cuBLAS libraries are host-provided and are not redistributed in these capsules. Their own installation terms apply. Diagnostic checkpoints and build completion do not establish model acceptance or constitute a production model release.

Downloads last month
-
GGUF
Model size
2B params
Architecture
granite_speech
Hardware compatibility
Log In to add your hardware

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for cstr/crispasr-granite5-validation-public

Quantized
(5)
this model