README: user-facing model card (files, run with codec.cpp tts-cli, license as upstream)

README.md

@@ -1,23 +1,41 @@
 ---
-license: mit
+license: other
+license_name: per-codec
+license_link: https://huggingface.co/BricksDisplay/codec.cpp-gguf#license
+tags:
+- codec.cpp
+- gguf
+- audio
+- neural-audio-codec
 ---
 
 # codec.cpp GGUFs
 
-Audio codec GGUFs for [codec.cpp](https://github.com/mybigday/codec.cpp): IBM `DAC.speech.v1.0`
-(24 kHz), NVIDIA NeMo Nano Codec (22.05 kHz, 0.6 kbps, 12.5 fps) and WavTokenizer-large-speech
-(75 tokens/s). Each comes as F32, F16 (no suffix) and Q8_0 / Q5_K_M / Q4_K_M.
+Neural audio codecs for [codec.cpp](https://github.com/mybigday/codec.cpp): turn
+speech into discrete tokens and back, offline on CPU or GPU.
 
-codec.cpp computes every weight at the precision stored in the GGUF: F16 files run F16 kernels,
-Q8_0 files run Q8_0 kernels. F16 is the default; the quantized files trade quality for size and
-speed (for DAC.speech, decode-only SNR vs F32 is ~37 dB at Q8_0, ~27 dB at Q5_K_M, ~23 dB at
-Q4_K_M on a speech clip; end-to-end encode+decode is lower because quantized encoders flip codes).
+| Codec | Sample rate | Token rate | Files |
+|---|---|---|---|
+| [IBM DAC.speech v1.0](https://huggingface.co/ibm-research/DAC.speech.v1.0) | 24 kHz | 75 frames/s | `ibm-research--DAC.speech*.gguf` |
+| [NVIDIA NeMo Nano Codec](https://huggingface.co/nvidia/nemo-nano-codec-22khz-0.6kbps-12.5fps) | 22.05 kHz | 12.5 frames/s (0.6 kbps) | `nemo-nano-codec-22khz-0.6kbps-12.5fps*.gguf` |
+| [WavTokenizer large (speech)](https://huggingface.co/novateur/WavTokenizer-large-speech-75token) | 24 kHz | 75 tokens/s | `wavtokenizer-large-speech-75tokens*.gguf` |
 
-## Changelog
+Each codec comes in several precisions: the file without a suffix is F16
+(recommended), `-F32` is full precision, `-Q8_0` / `-Q5_K_M` / `-Q4_K_M` are
+smaller and faster at some quality cost.
 
-- 2026-10-18: `ibm-research--DAC.speech-{Q8_0,Q5_K_M,Q4_K_M}.gguf` regenerated. Their quantized
-  conv kernels now use the layouts codec.cpp can run quantized: conv1d kernels flattened to
-  `(K*C_in, 1, C_out)`, ConvTranspose kernels `(C_in, C_out, K)`. The previous files used the
-  `(C_out, C_in, K)` layout, which only ran because older runtimes dequantized every weight to F32.
-  **These files need codec.cpp 961162a or newer.** F32/F16 files and the NeMo / WavTokenizer files
-  are unchanged.
+## Run
+
+```bash
+hf download BricksDisplay/codec.cpp-gguf ibm-research--DAC.speech.gguf --local-dir codecs
+
+./build/codec-cli encode --model codecs/ibm-research--DAC.speech.gguf --in speech.wav --out codes.npy
+./build/codec-cli decode --model codecs/ibm-research--DAC.speech.gguf --codes codes.npy --out decoded.wav
+```
+
+Build instructions, GPU flags and the Python API are in the
+[codec.cpp README](https://github.com/mybigday/codec.cpp).
+
+## License
+
+Each codec keeps its original license: DAC.speech CDLA-Permissive-2.0, NeMo Nano Codec the NVIDIA Open Model License, WavTokenizer MIT.