What AI can a NVIDIA DGX Spark 128 GB run?

Quick verdict

NVIDIA DGX Spark, 128 GB can run 29 of the 31 models tracked here at a comfortable speed, and another 1 slowly. The largest chat or coding model that fits is Qwen 3.5 122B A10B, which needs about 85.5 GB and generates roughly 12 words a second. The quickest is Gemma 4 E2B at about 62 words a second, which feels faster than you can read.

What limits this machine

Machines like this share fast soldered memory between the processor and the GPU, and the BIOS decides how much the GPU may use. About 115 GB of the 128 GB is usable here, at 273 GB/s. That makes large models affordable compared with a high-end graphics card, though the memory is slower than Apple's Max-class chips.

What it is good for

Chat and coding

Best chat pick: Qwen 3.5 122B A10B at about 12 words a second. Best coding pick: Qwen 3.6 35B A3B at about 35.

Images

• Stable Diffusion XL: runs great, about about 11 s per image
• FLUX.2 klein 4B: runs great, about about 19 s per image
• Z-Image Turbo 6B: runs great, about about 27 s per image

Video

• LTX-2.3: runs great, about about 2 min per 4 second clip
• Wan 2.2 5B: runs great, about about 7 min per 4 second clip
• Wan 2.2 14B: runs great, about about 14 min per 4 second clip

Voice and music

• Whisper large-v3 turbo: runs great, about about 42 s per 10 minutes of audio
• Kokoro TTS: runs great, about about 6 s per minute of speech
• Qwen3 TTS 1.7B: runs great, about about 30 s per minute of speech
• ACE-Step 1.5: runs great, about about 1 min per 1 minute song

Fastest models here

• Gemma 4 E2B: about 62 words a second, faster than you can read
• Gemma 4 E4B: about 38 words a second, smooth
• Qwen 3.5 4B: about 38 words a second, smooth
• Qwen 3.5 35B A3B: about 37 words a second, smooth
• Qwen 3.6 35B A3B: about 35 words a second, smooth

Memory headroom

Spare memory after loading the model decides how long a conversation or document can be before things slow down. With about 115 GB usable here:

• Qwen 3.5 122B A10B needs 85.5 GB, leaving 29.7 GB spare, plenty for long documents.
• gpt-oss 120B needs 68.8 GB, leaving 46.4 GB spare, plenty for long documents.
• Qwen 3.6 35B A3B needs 24.6 GB, leaving 90.6 GB spare, plenty for long documents.
• Qwen 3.5 35B A3B needs 23.6 GB, leaving 91.6 GB spare, plenty for long documents.
• Gemma 4 31B needs 21.5 GB, leaving 93.7 GB spare, plenty for long documents.

What does not fit, and by how much

• DeepSeek V3 671B needs about 424.7 GB and this machine tops out near 115 GB, so it is about 310 GB short.

How it compares

• Ryzen AI Max+ 395 (Strix Halo), 64 GB: 4 fewer models; drops Gemma 4 31B, Qwen 3.5 122B A10B, gpt-oss 120B and more.
• Ryzen AI Max+ 395 (Strix Halo), 128 GB: 2 fewer models; drops Gemma 4 31B, DeepSeek R1 32B.

Is an upgrade worth it?

The biggest single jump is Mac with M3 Ultra, 256 GB, which would unlock 1 more models, including Llama 3.3 70B.

Questions