How fast is Qwen3-4B on Mac Max?

The fastest of 3 measured configs is 186 tok/s single-stream (MLX-4BIT, mlx). The slowest is 55.9 tok/s.

Which quant of Qwen3-4B is fastest on Mac Max?

MLX-4BIT at 186 tok/s, ahead of Q4_K_M at 159.7, BF16 at 55.9. Plain decoding only.

Which backend is fastest for Qwen3-4B on Mac Max?

mlx at 186 tok/s, then llama.cpp at 159.7, xwen at 55.9. Each backend figure is its best plain-decoding row and may use a different quant.

Configs › Qwen3-4B › Qwen3-4B on Mac Max: 186 tok/s, 3 measured configs

Qwen3-4B on Mac Max: 186 tok/s, 3 measured configs

Measured Qwen3-4B speeds on Mac Max: fastest 186 tok/s single-stream (MLX-4BIT, mlx), 3 configs with quant, backend and context, each linked to its source, snapshot 2026-10-29.

Snapshot 2026-10-29 · 413 configs on AI boxes and 288 on GPU cards · 121 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Fastest measured single-stream run of Qwen3-4B on Mac Max: 186 tok/s (MLX-4BIT, mlx), snapshot 2026-10-29.

Every measured single-stream run of Qwen3-4B on Mac Max

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
186	MLX-4BIT	mlx	512	Extracted, source-linked	emihiggins	2026-09-29	source
159.7	Q4_K_M	llama.cpp	512	Extracted, source-linked	emihiggins	2026-09-29	source
55.9	BF16	xwen	3890	Extracted, source-linked	0rvar	2026-09-25	source

Best quant on Mac Max

Fastest plain-decoding run per quant, versus the fastest quant.

MLX-4BIT: 186 tok/s, fastest (mlx)
Q4_K_M: 159.7 tok/s, 14% lower (llama.cpp)
BF16: 55.9 tok/s, 70% lower (xwen)

Backends on Mac Max

Fastest plain-decoding run per backend. Quants can differ between rows.

mlx: 186 tok/s (MLX-4BIT)
llama.cpp: 159.7 tok/s (Q4_K_M)
xwen: 55.9 tok/s (BF16)

Questions and answers

How fast is Qwen3-4B on Mac Max?: The fastest of 3 measured configs is 186 tok/s single-stream (MLX-4BIT, mlx). The slowest is 55.9 tok/s.
Which quant of Qwen3-4B is fastest on Mac Max?: MLX-4BIT at 186 tok/s, ahead of Q4_K_M at 159.7, BF16 at 55.9. Plain decoding only.
Which backend is fastest for Qwen3-4B on Mac Max?: mlx at 186 tok/s, then llama.cpp at 159.7, xwen at 55.9. Each backend figure is its best plain-decoding row and may use a different quant.

See also: Qwen3-4B on every box · Mac Max hardware page (from S$3,499 in Singapore today).

Last updated: snapshot 2026-10-29; newest source check 2026-09-29.

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.