Qwen
- Qwen3 8B8B
Dense 8B from the Qwen3 generation.
- Qwen3 14B14B
Dense 14B. Q4 usually fits a 16 GB card; Q8 is tight there.
- Qwen3 30B-A3B30B
MoE: about 30B stored, about 3B active. Memory follows the stored count.
- Qwen3 235B-A22B235B
Large MoE. All experts sit in memory, so Q4 is still a big allocation.
- Qwen3.8 27B27B
Dense 27B vision-language model. The most-measured model in the current RigBase import.
- Qwen3.6 35B-A3B35B
MoE with 35B stored and 3B active. Faster than the parameter total suggests; memory still holds every expert.