Qwen3.5-9B-oQ6-mtp

Qwen3.5-9B-oQ6-mtp

This repository contains an MLX/oMLX-compatible quantization of Qwen/Qwen3.5-9B. It is a derivative quantized release; all credit for the original model architecture, training, and base release belongs to Qwen.

Summary

• Base model: Qwen/Qwen3.5-9B
• Quantization: oMLX oQ6 mixed-precision affine quantization, group size 64, bfloat16 conversion path, MTP preserved, VLM/vision tensors retained
• Effective bpw estimate from oMLX: 7.05
• Runtime target: oMLX / MLX on Apple Silicon
• Preserved features: MTP tensors, vision/VLM tensors, tokenizer, chat template
• License: Apache-2.0; see LICENSE

Verification

Local verification before release:

• safetensors metadata: {"format": "mlx"}
• safetensors total bytes: 8,505,251,264
• tensor count: 1289
• MTP tensor count: 29
• vision tensor count: 333
• oMLX local load test: passed
• text smoke test: passed (OK)
• image smoke test: passed (Red.)

See PUBLISH_CHECK.md for release verification details.

Files

README.md
LICENSE
PUBLISH_CHECK.md
chat_template.jinja
config.json
merges.txt
model-00001-of-00002.safetensors
model-00002-of-00002.safetensors
model.safetensors.index.json
preprocessor_config.json
tokenizer.json
tokenizer_config.json
vocab.json

Notes

This model was produced with oMLX oQ quantization and is intended for MLX/oMLX-compatible runtimes. No runtime sampling parameters are prescribed by this repository; choose inference settings appropriate for your application.