Update README.md

Author: subspecs

README.md

@@ -10,7 +10,26 @@ What features are currently implemented:
 - This library is compiled against the netstandard 2.1, that means you can use it both in .NET 5+ applications and the Unity game engine, allowing this to be used for games as well.
 
 You still have to build [s2.dll](https://github.com/rodrigomatta/s2.cpp) manually along with the ggml.
-Occasionally I'll include pre-built versions of S2 along with the ggml dll's, but usually they'll only come with CUDA/VULKAN/CPU support since I don't use crapple(apple).
+
+Occasionally I'll include pre-built versions of S2 along with the ggml dll's in [Releases](https://github.com/subspecs/FishS2Sharp/releases/), but usually they'll only come with CUDA/VULKAN/CPU support since I don't use crapple(apple).
+
+
+## Model variants
+
+GGUF files are available at [rodrigomt/s2-pro-gguf](https://huggingface.co/rodrigomt/s2-pro-gguf) on Hugging Face.
+
+| File | Size | Notes |
+|---|---|---|
+| `s2-pro-f16.gguf` | 9.9 GB | Full precision — reference quality |
+| `s2-pro-q8_0.gguf` | 5.6 GB | Near-lossless — recommended for 8+ GB VRAM |
+| `s2-pro-q6_k.gguf` | 4.5 GB | Good quality/size balance — recommended for 6+ GB VRAM |
+| `s2-pro-q5_k_m.gguf` | 4.0 GB | Smaller with still-good quality |
+| `s2-pro-q4_k_m.gguf` | 3.6 GB | Best compact variant so far in quick RU validation |
+| `s2-pro-q3_k.gguf` | 3.0 GB | Usable, but starts stretching short words |
+| `s2-pro-q2_k.gguf` | 2.6 GB | Lowest-size experimental variant |
+
+All variants include both the transformer weights and the audio codec in a single file.
+The quantized variants above were regenerated with the codec tensors (`c.*`) kept in `F16`, so only the AR transformer is quantized.
 
 Example usage:
 ```C#