Alibaba

Qwen3

reasoningMoE

Pick a size: the video memory shown is the Q4_K_M file (the protocol's quantisation) plus about 1.5 GB for a 4,000-token context.

0.6B

8 GB card

Parameters: 0,8B
Q4_K_M file: 0,5 Go
Recommended VRAM: ≥ 2,0 Go
Max context: 40 960

Measured on 1 card · 728 tok/s (RTX 5070 Ti)

1.7B

8 GB card

Parameters: 2B
Q4_K_M file: 1,2 Go
Recommended VRAM: ≥ 2,7 Go
Max context: 40 960

Measured on 1 card · 456 tok/s (RTX 5070 Ti)

8 GB card

Parameters: 4B
Q4_K_M file: 2,3 Go
Recommended VRAM: ≥ 3,8 Go
Max context: 40 960

Measured on 1 card · 225 tok/s (RTX 5070 Ti)

8 GB card

Parameters: 8,2B
Q4_K_M file: 4,7 Go
Recommended VRAM: ≥ 6,2 Go
Max context: 40 960

Measured on 1 card · 146 tok/s (RTX 5070 Ti)

14B

12 GB card

Parameters: 14,8B
Q4_K_M file: 8,4 Go
Recommended VRAM: ≥ 9,9 Go
Max context: 40 960

Measured on 1 card · 84,8 tok/s (RTX 5070 Ti)

32B

24 GB card

Parameters: 32,8B
Q4_K_M file: 18,4 Go
Recommended VRAM: ≥ 19,9 Go
Max context: 40 960

Not measured yet

30B-A3B

24 GB card

Parameters: 30,5B · 3B active
Q4_K_M file: 17,4 Go
Recommended VRAM: ≥ 18,9 Go
Max context: 40 960

Not measured yet

235B-A22B

> 32 GB

Parameters: 235,1B · 22B active
Q4_K_M file: 132,4 Go
Recommended VRAM: ≥ 133,9 Go
Max context: 40 960

Not measured yet

MoE (“30B-A3B”): all the memory of a 30B, but the speed of a model with 3B active parameters. “> 32 GB”: several cards, or part of the model in system memory (much slower).