Qwen2

Grouped-query attention with qkv bias. The architecture behind the Qwen2 and Qwen2.5 families.

Language 3 supported combinations Apple Silicon

Stacked slabs of dark slate, two of the gaps between them lit from within.

Qwen2 on this site

3 supported combinations across Apple Silicon.

Image rendered by Veizik for NVIDIA, development build (linux x86_64) · SD3.5-medium · NVIDIA RTX 3090 · Powered by Stability AI

Families

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.

Qwen2.5

Run Qwen2.5 directly from supported safetensors checkpoints with no converted model copy.

Variants

Every supported combination

Model	Model identifier	Short name	Checkpoint on disk	Veizik device memory	Context	Hardware
Qwen2.5-0.5B	qwen2.5-0.5b	0.5b	942 MiB	328 MiB	32,768	Apple Silicon
Qwen2.5-1.5B	qwen2.5-1.5b	1.5b	2,944 MiB	884 MiB	32,768	Apple Silicon
Qwen2.5-7B	qwen2.5-7b	7b	14,525 MiB	3,923 MiB	32,768	Apple Silicon

Veizik prepares supported weights in memory; no converted checkpoint is written to disk.

Elsewhere

Full catalogue

Filter every model and hardware combination Veizik supports.

Measurements

Every published result, with its device, protocol and date.

Documentation

Install, run, serve, and check an installation.