Qwen3
Adds per-head query and key normalisation and drops the qkv bias relative to Qwen2.
Language
1 supported combinations
Apple Silicon
Model families on this architecture
A family names one architecture. Support is declared per model and per target, so the same architecture can be open on one backend and absent from another.
Every supported combination
| Model | Command | Checkpoint on disk | Veizik device memory | Context | Hardware |
|---|---|---|---|---|---|
| Qwen3-1.7B | 1.7b | 3,281 MiB | 923 MiB | 32,768 | Apple Silicon |
Veizik prepares supported weights in memory; no converted checkpoint is written to disk.
How support for this architecture was checked
- Against a trusted reference implementation; argmax identical at every evaluated position. Maximum absolute logit difference 4.3e-05, measured 2026-09-16.