CLAP2 Medium AudioXL β€” GGUF

F32 and F16 GGUF conversions of laion/clap2-medium-audioxl-s1B for clap.cpp.

  • Joint text/audio embeddings: 640 dimensions, L2-normalized.
  • Audio frontend: 32 kHz; clap.cpp accepts WAV, MP3 and FLAC and resamples audio as needed.
  • Both encoders and tokenizer/frontend configuration are embedded in each single GGUF file.
  • F16 stores 2D weight matrices in IEEE FP16; other tensors and native activations remain F32.
  • CLAP2 block quantization is not supported by this release.

Use

Use clap.cpp commit 60def51cff6564eb35ef859487b1071d04e0ed52 or a compatible later version:

git clone https://github.com/gorse-io/clap.cpp
cd clap.cpp
git checkout 60def51cff6564eb35ef859487b1071d04e0ed52
git submodule update --init --recursive
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --target clap-cli -j2

hf download gorse-io/clap2-medium-audioxl-s1B-GGUF clap2-medium-audioxl-s1B-f16.gguf --local-dir models
./build/clap embedding --model models/clap2-medium-audioxl-s1B-f16.gguf --text 'Sound of a dog'
./build/clap embedding --model models/clap2-medium-audioxl-s1B-f16.gguf --audio dog.wav

Substitute f32 for f16 to use F32 storage. These files target clap.cpp, not llama.cpp.

Provenance and validation

Source revision: 3a8d9d79cb98e4d409fce24092a69718a821a7ad. Conversion commit: 58ca68f4ea085a20d052a5f78666072819067f5a, whose tree is identical to the merged public commit 60def51cff6564eb35ef859487b1071d04e0ed52. The exact source-checkpoint SHA-256, storage audit, runtime identity, official-reference fixtures and per-format native text/audio validation are included in this repository.

Both formats were checked against the pinned official CLAP2 implementation. F16 comparisons use identically storage-rounded reference weights, not the untouched F32 weights. These deterministic probes establish implementation agreement, not retrieval-quality benchmarks.

License and attribution

Original model weights: LAION, declared Apache-2.0 in the pinned upstream model card. See the original model card. clap.cpp is separately licensed under Apache-2.0. This conversion does not change the source-model license.

Downloads last month
116
GGUF
Model size
0.2B params
Architecture
clap2
Hardware compatibility
Log In to add your hardware

16-bit

32-bit

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

Model tree for gorse-io/clap2-medium-audioxl-s1B-GGUF

Quantized
(1)
this model

Collection including gorse-io/clap2-medium-audioxl-s1B-GGUF