Intel · 128 GB HBM2e · 3276 GB/s bandwidth ·
Mistral · 23.6B · model ctx up to 128K
Yes — the Gaudi 3 128GB (128 GB HBM2e) can run Devstral-Small-2507. At Q4 with the default 8K context it needs ≈ 18.7 GB VRAM (weights 12.98 GB + KV cache 4.72 GB + 1 GB runtime overhead), which fits within 128 GB. Decode at Q4/8K: ≈ 88.3 tok/s (estimated, batch 1). FP16 (≈ 52.92 GB) also fits, at ≈ 24.3 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
16.34 GB
Yes
88.3est.
Q4
8K default
18.7 GB
Yes
88.3est.
Q4
32K
32.86 GB
Yes
88.3est.
FP16
4K
50.56 GB
Yes
24.3est.
FP16
8K default
52.92 GB
Yes
24.3est.
FP16
32K
67.08 GB
Yes
24.3est.
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).