Llama 3.2 3B
3B
~33 tok/s per stream
DetailsHide details
Where the memory goes
- Model weights2.29 GB
- KV cache2.62 GB
- Activations0.01 GB
- CUDA & framework1.01 GB
- Precision
- INT4 / AWQ
- Serving preset
- 8K context × 4 users
- Users at this context
- 4
- Spare context in leftover VRAM
- 26K
- Fit score
- 42/100 · Very high