Apple · 96 GB LPDDR5 · 819 GB/s bandwidth ·
Google · 8.0B · model ctx up to 128K
Yes — the Mac Studio M3 Ultra 96GB (96 GB LPDDR5) can run gemma-4-E4B-it. At Q4 with the default 8K context it needs ≈ 7.0 GB VRAM (weights 4.4 GB + KV cache 1.6 GB + 1 GB runtime overhead), which fits within 96 GB. Decode at Q4/8K: ≈ 65.1 tok/s (estimated, batch 1). FP16 (≈ 18.6 GB) also fits, at ≈ 17.9 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
6.2 GB
Yes
65.1est.
Q4
8K default
7.0 GB
Yes
65.1est.
Q4
32K
11.8 GB
Yes
65.1est.
FP16
4K
17.8 GB
Yes
17.9est.
FP16
8K default
18.6 GB
Yes
17.9est.
FP16
32K
23.4 GB
Yes
17.9est.
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).