Clarify the runtime bit-width requirement and what is dropped from the export

README.md

@@ -78,8 +78,9 @@ quantized; they stay float. Quantized tensors are stored as MLX's
 | `int8/tokenizer_config.json` | 16.7 kB | Chat template and special tokens |
 
 Each variant is self-contained: one directory is everything needed to run it.
-The multi-token-prediction draft head (`mtp.*`) and the vision tower are not
-included — no runtime here reads them, and they were 14 MB of dead download.
+The vision tower and the multi-token-prediction draft head (`mtp.*`) are not
+included — no runtime here reads them. Dropping the draft head alone took 14 MB
+off the INT4 download compared with the previous revision.
 
 ## Measured quality
 
@@ -140,10 +141,11 @@ let response = try model.generate(
 ```
 
 > **INT5 and INT8 need a runtime that reads the bit width from `config.json`.**
-> Older speech-swift versions build every `QuantizedLinear` with `bits = 4`
-> hardcoded and will misread an INT5 or INT8 file — it loads without error and
-> produces garbage. Use a speech-swift version that takes `quantization_bits`
-> and `quantization_group_size` from `config.json`, or stay on INT4.
+> Older speech-swift versions build every `QuantizedLinear` and
+> `PreQuantizedEmbedding` with `bits = 4` hardcoded, which fits the INT4 file
+> only; they will not read an INT5 or INT8 file correctly. Use a speech-swift
+> version that takes `quantization_bits` and `quantization_group_size` from
+> `config.json`, or stay on INT4.
 
 ### Python