128 GB · 273 GB/s
What can the NVIDIA DGX Spark run?
A small box with 128 GB that runs 100B+ mixture-of-experts models.
Best models for each task
Our top three for this machine. Chat and image tasks assume an 8K-token conversation; coding, documents and agents assume 32K.
Chat & writing24 of 25 run
- Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.)Runs wellUD-IQ4_XS · needs 95.8 GB at 8K · 30–55tok/sFast
- Qwen3.8 27BRuns greatQ8_0 · needs 31.6 GB at 8K · 5.5–9tok/sReadable
- Gemma 4 26B-A4BRuns greatQ8_0 · needs 29.3 GB at 8K · 25–45tok/sSmooth
Coding25 of 26 run
- Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.)Runs wellUD-IQ4_XS · needs 96.4 GB at 32K · 25–50tok/sSmooth
- DeepSeek V4 Flash 0731Just fitsUD-IQ3_XXS · needs 105 GB at 32K · 20–40tok/sSmooth
- Qwen3.8 27BRuns greatUD-IQ3_S · needs 16.1 GB at 32K · 12–18tok/sSmooth
Long documents19 of 20 run
- Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.)Runs wellUD-IQ4_XS · needs 96.4 GB at 32K · 25–50tok/sSmooth
- DeepSeek V4 Flash 0731Just fitsUD-IQ3_XXS · needs 105 GB at 32K · 20–40tok/sSmooth
- Qwen3.8 27BRuns greatUD-IQ4_XS · needs 18.4 GB at 32K · 10–16tok/sReadable
Agents & tools25 of 26 run
- Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.)Runs wellUD-IQ4_XS · needs 96.4 GB at 32K · 25–50tok/sSmooth
- DeepSeek V4 Flash 0731Just fitsUD-IQ3_XXS · needs 105 GB at 32K · 20–40tok/sSmooth
- Qwen3.8 27BRuns greatUD-IQ3_S · needs 16.1 GB at 32K · 12–18tok/sSmooth
Images in12 of 13 run
- Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.)Runs wellUD-IQ4_XS · needs 95.8 GB at 8K · 30–55tok/sFast
- Qwen3.8 27BRuns greatQ8_0 · needs 31.6 GB at 8K · 5.5–9tok/sReadable
- Qwen3.6 35B-A3BRuns greatQ8_0 · needs 39.9 GB at 8K · 30–55tok/sFast
Make images6 of 6 run
- Stable Diffusion XL Base 1.0Runs greatFP16 · needs 7.9 GB
- FLUX.2 [klein] 4BRuns greatBF16 · needs 14.0 GB
- Z-Image-TurboRuns greatBF16 · needs 17.0 GB
Speak text7 of 7 run
- Qwen3-TTS 1.7BRuns greatBF16 · needs 5.5 GB
- Kokoro-82MRuns greatFP32 · needs 1.4 GB
- Chatterbox Multilingual V3Runs greatFull · needs 4.2 GB
Transcribe7 of 7 run
- Voxtral Mini 4B RealtimeRuns greatBF16 · needs 17.0 GB
- Whisper large-v3-turbo (via faster-whisper)Runs greatFP16 · needs 2.6 GB
- Nemotron 3.5 ASR Streaming 0.6BRuns greatFP32 · needs 3.5 GB
Biggest model it runs well
Qwen3.8 Flash Next 125B-A6B (+51B n-gram emb.) (UD-IQ4_XS, 93.7 GB file) at about 30–55tok/sFast.
Should you buy it?
Buy it if
You want 100B+ mixture-of-experts models on your desk, like Qwen3.8 Flash Next (what most Spark owners run now) or gpt-oss-120b.
Skip it if
Your models fit in 32 GB — an RTX 5090 is several times faster on them.
Watch out for
- Bandwidth (273 GB/s) is low: big dense models are slow.
- Runs Linux on ARM; some tools assume x86.
Measured on this hardware
Real runs from public sources, not estimates. All measurements.