Apple · 96 GB LPDDR5 · 819 GB/s bandwidth ·
Alibaba · 8.2B · model ctx up to 128K
Yes — the Mac Studio M3 Ultra 96GB (96 GB LPDDR5) can run Qwen3-8B. At Q4 with the default 8K context it needs ≈ 7.15 GB VRAM (weights 4.51 GB + KV cache 1.64 GB + 1 GB runtime overhead), which fits within 96 GB. Decode at Q4/8K: ≈ 63.6 tok/s (estimated, batch 1). FP16 (≈ 19.04 GB) also fits, at ≈ 17.5 tok/s (estimated). Estimates come from memory-bandwidth math; rows tagged measured override estimates. Relative ranking is more reliable than absolute tok/s. Methodology.
Fit & speed by quant and context
Quant
Context
VRAM needed
Fits?
tok/s (decode)
Q4
4K
6.33 GB
Yes
63.6est.
Q4
8K default
7.15 GB
Yes
63.6est.
Q4
32K
12.07 GB
Yes
63.6est.
FP16
4K
18.22 GB
Yes
17.5est.
FP16
8K default
19.04 GB
Yes
17.5est.
FP16
32K
23.96 GB
Yes
17.5est.
Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured
anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s
means the model is far beyond this card (offload-only territory).