Card: safetensors checkpoint for MLX and transformers

README.md

@@ -31,6 +31,7 @@ This model is not affiliated with TypeSafe, and it is not Jev.
 
 | File | What it is |
 | --- | --- |
+| `model-0000*-of-00005.safetensors`, `config.json`, tokenizer files | Merged model, BF16 safetensors (8.2 GB): for the MLX backend and transformers. Same weights as the BF16 GGUF, bit for bit |
 | `spark-x2.5-4b-rizzo-flow-lora-q8_0.gguf` | Merged model, Q8_0 (4.4 GB): the default for `rizzo serve` |
 | `spark-x2.5-4b-rizzo-flow-lora-bf16.gguf` | Merged model, BF16 (8.2 GB) |
 | `spark-x2.5-4b-rizzo-flow-lora-q4_k_m.gguf` | Merged model, Q4_K_M (2.6 GB), for GPUs with little memory |
@@ -57,7 +58,25 @@ uv run rizzo serve             # → http://127.0.0.1:8017/playground
 [`rizzoaiacademy/rizzo-flow-1.7b`](https://huggingface.co/rizzoaiacademy/rizzo-flow-1.7b)), and
 `--weights base` the original Spark-X2.5 GGUF files. The server answers as `rizzo-flow-4b-q8_0`.
 
-The files are plain GGUF (architecture `spark2_5`; Rizzo Flow runs them on llama.cpp release b11081), but the model is
+**MLX** (Apple Silicon, or MLX-CUDA) uses the safetensors checkpoint at the root of this
+repository instead of the GGUF files:
+
+```bash
+uv sync --locked --extra mlx
+uv run --no-sync rizzo download --backend mlx      # the safetensors only, 8.2 GB
+uv run --no-sync rizzo serve --backend mlx --bits 8
+```
+
+It answers as `rizzo-flow-4b-bf16` (or `…-q8`/`…-q4` with `--bits`). On an RTX 5060 Ti with MLX-CUDA, BF16, the
+4B gives the same prompts (same hashes) and the same probabilities as llama.cpp BF16 within 0.001.
+
+**transformers** loads the checkpoint with the original Spark code (`trust_remote_code=True`,
+`modeling_spark.py` is XHToken's file, unchanged): `AutoModelForCausalLM.from_pretrained(
+"rizzoaiacademy/rizzo-flow", trust_remote_code=True, dtype="bfloat16")`. Use the prompts of
+Rizzo Flow and read the answer-letter logits at the last position; generating text is not what
+it was trained for.
+
+The GGUF files are plain GGUF (architecture `spark2_5`; Rizzo Flow runs them on llama.cpp release b11081), but the model is
 meant to be **read, not chatted with**: Rizzo Flow builds the prompt, then reads the logits of the
 answer letters. In a chat app it behaves like a lightly changed Spark-X2.5, and the fine-tune
 brings nothing there.