Qwen3.5-0.8B
Fast proof point for compression behavior and tier quality.
Supported Models
Start with one family and make the result undeniable.
The first public model surface is centered on Qwen3.5 because it is easier to believe a strong compression story when the model line, the tiers, and the setup path all stay coherent.
Qwen3.5-2B
First model where smaller and still useful becomes a real product story.
Qwen3.5-4B
Mainstream pressure test for the launch surface.
Qwen3.5-9B
Where memory ceilings and multi-GPU costs start to bite hard.
The tier ladder
Base
Original fp16. This is the reference point everything else has to justify.
Parity
The safety-first tier for people who want aggressive size reduction without giving up the behavior they trust.
Frontier High / Low
The compression-first tiers for people chasing much smaller artifacts and willing to choose how hard they want to push.