Are there plans to release a 122B-A10B version?

[nickludlam]

Hi, the Qwen 3.5 122B model is one of my favourites to use locally, and I would love to know if the team has plans to release a similar sized MoE model for 3.6?

[EJRIV]

Same here! Is a mid sized qwen3.6 coming?

[lessu]

+1

[david-king-neuma]

+1

[bendelec]

Fully Agree!

There is currently a real gap in the model world for the growing amount of 128gb overall / 96gb vram unified memory machines (128gb Mac hardware, Strix Halo, ...).

I'd personally favor something like a 70-80b overall size, (let's say 72b-a9b), because I think this is where the real sweet spot is for my kind of hardware (128gb strix halo 395), but I'd also be very grateful for a 122b-a10b like OP suggests. It would need more compromise in quantization, but it should be doable.

Status quo:

• 70b-ish dense models are mostly several generations old and make you feel it. Dense models of that size also slow down a lot in the unified memory machines due to memory bandwidth limitations.
• New radically MoE models like qwen3-next 80b-a3b are fast, but in many tasks they fail to hide the shortcomings of their very small experts / active parameters. I've been struggling with this for weeks, dithering between qwen3.6-27b dense and qwen3-next 80b-a3b, and not being really happy with either.