Skip to content
localmodel.run

text model · phi · Linux

Can I run Phi-4-mini 3.8B on Nvidia DGX Spark (128GB)?

Compatibility verdict VRAM threshold engine
Yes, it runs GPU accelerated ~71 tok/s est.

Yes. Phi-4-mini 3.8B runs on Nvidia DGX Spark (128GB) at Q4_K_M (~3.8 GB of ~112 GB usable).

Needs ~3.8 GB Device usable ~112 GB

Runs at Q4_K_M using ~3.8 GB of ~112 GB usable. You have room for FP16 for higher quality.

That figure is at a 4k context and moves about ±15% as context length changes. Nvidia DGX Spark (128GB) leaves ~108.2 GB of headroom, room to step up to FP16 for higher quality.

Q4_K_M needed
~3.8 GB
Usable on device
~112 GB
Device memory
128 GB
Best quant
Q4_K_M
Share on X Share on Reddit

Which quant fits

Quant ladder vs ~112 GB usable
Q2_K
~2.9 GB
Q3_K_M
~3.2 GB
Q4_K_M
~3.8 GB
Q5_K_M
~4 GB
Q6_K
~4.4 GB
Q8_0
~5.4 GB
FP16
~8.9 GB
The line marks Nvidia DGX Spark (128GB)'s ~112 GB budget; rungs past it are too large.

Running cost · estimate

Power & economics est.
Power draw
~140 W
Electricity / 1M tokens
~$0.08
Pays for itself after
~9,521M tok

At ~$0.15/kWh and the estimated ~71 tok/s, a million generated tokens costs about $0.08 in electricity. Versus a hosted API at ~$0.5 per million tokens, the ~$3,999 Nvidia DGX Spark (128GB) pays for itself after roughly 9,521 million tokens, so local hardware is mostly a fixed cost, not a per-token one. TDP is the peak draw, so this is an upper bound. Assumptions.

Run it

Install commands Linux

Pick your tool. All 3 load the same Q4_K_M weights.

Ollama
$ ollama run phi4-mini:3.8b
llama.cpp
$ llama-cli -hf bartowski/microsoft_Phi-4-mini-instruct-GGUF:Q4_K_M
LM Studio
$ lms get bartowski/microsoft_Phi-4-mini-instruct-GGUF
Model phi
Parameters
3.8B
Q4_K_M size
2.49 GB
Q8_0 size
4.08 GB
Context
128k
Ollama tag
phi4-mini:3.8b

Set up this model at 4K →

Full Phi-4-mini 3.8B requirements →
Device Linux
Memory
128 GB unified
Usable for weights
~112 GB
Power draw
~140 W
Best runtime
llama.cpp (CUDA) / Ollama (CUDA)
Best models for Nvidia DGX Spark (128GB) →

You could also run

Run Phi-4-mini 3.8B on other hardware

Nvidia GeForce RTX 2060 (6GB)Apple M1 (8GB)8GB RAM Laptop (CPU/iGPU only)iPhone 15 ProiPhone 16iPhone 16 Pro

FAQ

Can Nvidia DGX Spark (128GB) run Phi-4-mini 3.8B?

Yes. Phi-4-mini 3.8B runs on Nvidia DGX Spark (128GB) at Q4_K_M (~3.8 GB of ~112 GB usable).

How much memory does Phi-4-mini 3.8B need?

Nvidia DGX Spark (128GB) has room to spare. At Q4_K_M the weights are ~2.49 GB; with KV cache and runtime overhead, budget ~3.8 GB at a 4k context.

What is the best tool to run Phi-4-mini 3.8B on Linux?

Ollama for a simple setup; vLLM for the most speed. vLLM shines for multi-user serving/throughput. For a single local chat, Ollama or llama.cpp is simpler and lighter.

Embed this

Phi-4-mini 3.8B on Nvidia DGX Spark (128GB) compatibility badge A live badge for your model card or README, updated as the data is.
Markdown
$ [![Phi-4-mini 3.8B on Nvidia DGX Spark (128GB)](https://localmodel.run/badge/phi-4-mini-3.8b/nvidia-dgx-spark-128gb.svg)](https://localmodel.run/can-i-run/phi-4-mini-3.8b/nvidia-dgx-spark-128gb)

Sources

Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.