netw52623/Qwen3.6-35B-A3B-MTP-Holo3-Qwopus-10bit-mlx

Quant of https://huggingface.co/nightmedia/Qwen3.6-35B-A3B-MTP-Holo3-Qwopus-BF16 which quantizes some layers to 8-bit and keeps a good portion in bf16, arriving at ~10 bpw "10 bit".

This is text-only and includes a jinja chat template that enforces no thinking (even with no thinking, model still thinks when problems get complicated, and skips useless interim thinking in the middle of research/exploration before sufficient context is acquired). Using text-only mlx-lm inferencing consistently provided higher quality research/reasoning so no visual support to decrease size a bit.

I found this specific quant offers some of the tighter thinking of full bf16 of this model, but also some of the wider/explorative nature of 8bit quant. As such I find it manages better overall research and double-checking thinking in agentic loops, offering better results than bf16 or 8bit after many hours of usage.

Recommended sampling params:

Temp: 0.6
Top P: 0.95
Top K: 20
Presence Penaly: 1.3

(Model does fail tool calls sometimes, so consistent usage requires a robust harness, but in personal use is 2x faster than standard Qwen3.6 35b at finishing most tasks, has better spatial understanding, and can solve more complex tasks consistently)

Big thanks to nightmedia for putting together a great model.