Can DGX Spark (GH200 480GB) run Hy3?

NVIDIA · 480 GB LPDDR5X · 500 GB/s bandwidth · Tencent · 295.0B (active 21.0B) · MoE · model ctx up to 256K

CUDA

Yes — the DGX Spark (GH200 480GB) (480 GB LPDDR5X) can run Hy3. At Q4 with the default 8K context it needs ≈ 167.45 GB VRAM (weights 162.25 GB + KV cache 4.2 GB + 1 GB runtime overhead), which fits within 480 GB. Decode at Q4/8K: ≈ 15.2 tok/s (estimated, batch 1). FP16 needs ≈ 595.2 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. Buy DGX Spark (GH200 480GB) on Amazon →

Fit & speed by quant and context

Quant	Context	VRAM needed	Fits?	tok/s (decode)
Q4	4K	165.35 GB	Yes	15.2 ± est.
Q4	8K default	167.45 GB	Yes	15.2 ± est.
Q4	32K	180.05 GB	Yes	15.2 ± est.
FP16	4K	593.1 GB	No	7.8 ± est.
FP16	8K default	595.2 GB	No	7.7 ± est.
FP16	32K	607.8 GB	No	7.4 ± est.

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	DGX Spark (GH200 480GB) VRAM
Q4	162.25 GB	4.2 GB	1 GB	167.45 GB	480 GB
FP16	590.0 GB	4.2 GB	1 GB	595.2 GB	480 GB

Other GPUs that run Hy3

• B200 192GB SXM — ≈242.4 tok/s Q4 est.
• MI300X 192GB — ≈160.6 tok/s Q4 est.
• Mac Studio M2 Ultra 192GB — ≈24.2 tok/s Q4 est.

All GPUs for Hy3 →

Other models for the DGX Spark (GH200 480GB)

• MiniCPM5-1B — ≈294.6 tok/s Q4 est.
• Qwen3.5-2B — ≈159.1 tok/s Q4 est.
• DeepSeek-R1-0528-Qwen3-8B-layer-mix-bpw-3.8-mlx — ≈138.3 tok/s Q4 est.
• Qwen3.6-35B-A3B — ≈106.1 tok/s Q4 est.
• Qwen3.5-35B-A3B — ≈106.1 tok/s Q4 est.
• GLM-4.7-Flash — ≈106.1 tok/s Q4 est.
• Qwen3-30B-A3B-Instruct-2507 — ≈106.1 tok/s Q4 est.
• Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 — ≈106.1 tok/s Q4 est.

All models on the DGX Spark (GH200 480GB) →