Hardware · SoC / APU · NVIDIA

NVIDIA GB10 Grace Blackwell

Shares one pool of LPDDR5X memory between CPU and GPU at 273 GB/s. What fits depends on how much memory the system has.

What it runs

Largest models without offloading on the GB10 mini workstation (128 GB), 8K context.

Variant
Llama 3.1 70B Instruct
70.6B · Q4_K_M via llama.cpp
Runs well
~4 tok/s est.
Variant
Llama 3.3 70B Instruct
70.6B · Q4_K_M via llama.cpp
Runs well
~4 tok/s est.
Variant
Mixtral 8x7B Instruct v0.1
46.7B · BF16 via SGLang
Runs well
~7 tok/s est.
Variant
Qwen2.5 32B Instruct
32.8B · BF16 via SGLang
Runs well
~3 tok/s est.
Variant
Qwen2.5-Coder 32B Instruct
32.8B · Q8_0 via llama.cpp
Runs well
~5 tok/s est.
Variant
DeepSeek-R1-Distill-Qwen-32B
32.8B · BF16 via SGLang
Runs well
~3 tok/s est.
Variant
Qwen3 30B-A3B
30.5B · Q8_0 via llama.cpp
Runs well
~44 tok/s est.
Variant
Gemma 3 27B IT
27.4B · BF16 via SGLang
Runs well
~3 tok/s est.
Variant
Gemma 2 27B IT
27.2B · BF16 via SGLang
Runs well
~3 tok/s est.
Variant
Mistral Small 24B Instruct 2501
23.6B · BF16 via SGLang
Runs well
~4 tok/s est.

System GB10 mini workstation (128 GB)
Runs models up to ~160B parameters at 4-bit in unified memory. What runs

Throughput on systems containing this device.

No performance measurements recorded yet.

Contributions are not open yet, so there is nothing here from members.

Contributions are not open yet, so there is nothing here from members.

No source records linked.