qwen-3.5-0.8b-Hedgehog-MLX-fp16
This is the MLX FP16 export of qwen-3.5-0.8b-Hedgehog, a Qwen 3.5 0.8B
SimpleRAG document extractor. It is also the source checkpoint for the matching
oMLX oQ4/oQ6/oQ8 and oQ4e/oQ6e/oQ8e builds.
Compatibility repair
The merged Hugging Face checkpoint identifies itself as qwen3_5_text, while
current mlx-lm provides the architecture module as qwen3_5. This export uses
the loader-compatible name and applies mlx-lm's Qwen3.5 sanitation contract:
model.language_model.*becomeslanguage_model.model.*;- Hugging Face depthwise Conv1d weights are moved from
[out, 1, kernel]to MLX's[out, kernel, 1]layout; - Qwen3.5 RMSNorm weights receive the required one-time
+1conversion; - ordinary floating tensors are FP16;
A_logremains 16-bit BF16, matching mlx-lm's architecture-specific cast predicate. - stale MTP declarations are removed because this checkpoint contains no MTP tensors.
Build the oMLX variants on Apple Silicon
Run build_omlx_variants.py from an environment that can import the omlx
package. By default it creates exactly these six sibling directories and runs
an extraction smoke test after every build:
qwen-3.5-0.8b-Hedgehog-MLX-oQ4qwen-3.5-0.8b-Hedgehog-MLX-oQ6qwen-3.5-0.8b-Hedgehog-MLX-oQ8qwen-3.5-0.8b-Hedgehog-MLX-oQ4eqwen-3.5-0.8b-Hedgehog-MLX-oQ6eqwen-3.5-0.8b-Hedgehog-MLX-oQ8e
/Applications/oMLX.app/Contents/MacOS/python3 build_omlx_variants.py --source . --output-root ..
No standard MLX affine model is requested or generated. oQe reuses its activation-importance cache across the three enhanced depths.
Intended inference contract
Use temperature 0, disable thinking, supply the retrieved document evidence, and state the exact JSON leaf paths and scalar types required by SimpleRAG.
This package contains the accepted r5 merged weights. Release evaluation of the matching Q8 GGUF scored 71 on BenchLocal DataExtract-15 (88 in the Complex Documents category), versus 72 and 94 for the prior r3 release. Its 24-case thinking suite detected no repetition loops and no length-cap runaways; 23 of 24 cases produced a final answer. The known extraction weaknesses are multi-entity records and synthetic complex invoices.
The Windows exporter validates every converted tensor exactly against the merged Hugging Face source. MLX runtime inference and every oQ/oQe variant must still be smoke-tested on Apple Silicon; conversion does not guarantee that a quantized variant retains the source model's score.