oQe uses an imatrix sensitivity calibration to assign precision selectively, rather than applying a uniform bit-width to every tensor. The generated oq_imatrix_report.json records the resulting quantization allocation.

python run_oqe_conversion.py Ornith-1.5-9B-bf16 --models-dir /path/to/models --oq-level 6.0 --imatrix-samples 128 --imatrix-seq-length 512 --mtp-donor /path/to/Qwen3.5-9B-bf16 --mtp-source-repo https://huggingface.co/Qwen/Qwen3.5-9B --source-repo https://huggingface.co/PleIAs/Ornith-1.5-9B

Lightning MTP

The donor MTP head was injected into the staged source before conversion, so it was included in the normal MTP sanitize, imatrix, and quantization passes.

Donor checkpoint: Qwen3.5-9B-bf16
Donor repository / credit: https://huggingface.co/Qwen/Qwen3.5-9B

Files

model.safetensors or model-*-of-*.safetensors: quantized MLX weights
model.safetensors.index.json: shard index (present for sharded outputs)
config.json: model configuration
oq_imatrix_report.json: imatrix calibration and quantization report

djrsystemservices/Ornith-1.5-9B-oQ6e-mtp

Ornith-1.5-9B-oQ6e-mtp-corrected

Source and credit

Conversion

Lightning MTP

Files

djrsystemservices/
Ornith-1.5-9B-oQ6e-mtp