Can the DGX Spark (GB10) run Gemma 4 31B?
Yes. Gemma 4 31B fits the DGX Spark (GB10)'s 128 GB at Q4_K_M (about 20.8 GB needed), running at an estimated 8.3 tok/s. There's room for up to FP16 if you want higher quality.
Every quantization
Gemma 4 31B (32.7B) on 128 GB
| Quant | Weights | Needs | Fit | Est. speed |
|---|---|---|---|---|
| Q4_K_M | 19 GB | 20.8 GB | Fits easily | ~8.3 tok/s |
| Q5_K_M | 23 GB | 24.8 GB | Fits easily | ~6.8 tok/s |
| Q8_0 | 35 GB | 36.8 GB | Fits easily | ~4.5 tok/s |
| FP16 | 65 GB | 66.8 GB | Fits easily | ~2.4 tok/s |
Estimates. Needs = weights + 1.8 GB for KV cache and runtime at modest context. Speed is single-stream decode scaled from the DGX Spark (GB10)'s 273.2 GB/s bandwidth (Estimated from memory bandwidth (llama.cpp decode)). Long contexts need more memory and run slower.
Run it
Copy, paste, go
llama-server -hf google/gemma-4-31B-it-qat-q4_0-gguf -ngl 99Ollama pulls a Q4 build by default. Pick a larger quant only if the table above says it fits. Full Gemma 4 31B details →
Compare
Same model, other hardware - and other models on this card
Gemma 4 31B on other hardware
- RTX 5090~54 tok/s
- RTX 5090 D~54 tok/s
- RTX 6000D~47 tok/s
- RTX 4090~33 tok/s
- RTX 3090~29 tok/s
- Radeon RX 7900 XTX~29 tok/s
- MacBook Pro 16" M4 Max~16 tok/s
- Arc Pro B60~14 tok/s
Also runs on the DGX Spark (GB10)
- gpt-oss-120b~35 tok/s
- Qwen2.5 72B~3.9 tok/s
- Llama 3.3 70B~3.9 tok/s
- Qwen3.6 35B A3B~60 tok/s
- Qwen3.5 35B A3B~60 tok/s
- dolphin-2.9.1-yi-1.5-34b~7.8 tok/s
- Qwen3 32B~8.3 tok/s
- GLM-4.7-Flash~60 tok/s
Frequently asked
Quick answers
Can the NVIDIA DGX Spark (GB10) run Gemma 4 31B?
Yes. Gemma 4 31B fits the DGX Spark (GB10)'s 128 GB at Q4_K_M (about 20.8 GB needed), running at an estimated 8.3 tok/s. There's room for up to FP16 if you want higher quality.
How much VRAM does Gemma 4 31B need?
About 19 GB of weights at Q4_K_M, 35 GB at Q8_0 and 65 GB at FP16. Add roughly 1.8 GB for the KV cache and runtime at modest context lengths.
How fast is Gemma 4 31B on the DGX Spark (GB10)?
Roughly 8.3 tokens per second at Q4_K_M (single user, llama.cpp-style decode). This is an estimate from the DGX Spark (GB10)'s 273.2 GB/s memory bandwidth.