README: precision notes and changelog

README.md

@@ -1,3 +1,23 @@
----
-license: mit
----
+---
+license: mit
+---
+
+# 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.
+
+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).
+
+## Changelog
+
+- 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.