Can Gaudi 3 128GB run DeepSeek-V3.2?

Intel · 128 GB HBM2e · 3276 GB/s bandwidth · DeepSeek · 685.4B (active 37B) · MoE · model ctx up to 128K

No — DeepSeek-V3.2 at Q4 needs ≈ 381.67 GB even at 4K context, beyond the Gaudi 3 128GB's 128 GB HBM2e. It would only run with heavy CPU/disk offload. Decode at Q4/8K: ≈ 17.8 tok/s (estimated, batch 1). FP16 needs ≈ 1379.2 GB — does not fit on this card. 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	381.67 GB	No	18.1 est.
Q4	8K default	385.37 GB	No	17.8 est.
Q4	32K	407.57 GB	No	15.9 est.
FP16	4K	1375.5 GB	No	—
FP16	8K default	1379.2 GB	No	—
FP16	32K	1401.4 GB	No	—

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).

VRAM breakdown at 8K context

Quant	Weights	KV cache	Overhead	Total needed	Gaudi 3 128GB VRAM
Q4	376.97 GB	7.4 GB	1 GB	385.37 GB	128 GB
FP16	1370.8 GB	7.4 GB	1 GB	1379.2 GB	128 GB

Other GPUs that run DeepSeek-V3.2

No catalog GPU fully fits DeepSeek-V3.2 at Q4/8K — see the model page for offload estimates.

All GPUs for DeepSeek-V3.2 →

Other models for the Gaudi 3 128GB

• Qwen3-0.6B — ≈2605.9 tok/s Q4 est.
• gemma-3-1b-it — ≈2084.7 tok/s Q4 est.
• Llama-3.2-1B-Instruct — ≈1737.3 tok/s Q4 est.
• Qwen3-1.7B — ≈1042.4 tok/s Q4 est.
• Qwen3-30B-A3B — ≈694.9 tok/s Q4 est.
• Qwen3-Coder-30B-A3B-Instruct — ≈694.9 tok/s Q4 est.
• GLM-4.7-Flash — ≈694.9 tok/s Q4 est.
• Qwen3-30B-A3B-Instruct-2507 — ≈694.9 tok/s Q4 est.

All models on the Gaudi 3 128GB →