Skip to content

Hy3 on DGX Spark: VRAM, speed and fit

Hy3 at 2-bit runs on the DGX Spark 128 GB with 32K of context at about 22 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on DGX Spark 128 GB · 2-bit (Q2_K) · Ollama / LM Studio · Chat

GPU memory
121 GB
of 128 GB
Context
32K
auto
Speed
22 tok/s
9.8 when full
First reply
4.8 s
2K prompt

Where the memory goes

Model
103 GB
Context
10.0 GB
Overhead
620 MB
OS
7.7 GB
Free
6.7 GB
Memory
121 GB of 128 GB
Context
32K tokens · auto
Speed
22 tok/s · 9.8 when full
First reply
4.8 s for a 2K prompt
Running cost
$3.15/mo at home · $287/mo on a cloud 4× A40

Validation: Validated with conditions

What Hy3 on DGX Spark was checked against
CheckTargetResultStatus
Fits in GPU memory120 GB usable114 GB neededPass
Headroom5% free6.7 GB free (6%)Pass
Context for chat16K or more32KPass
Writing speed15 tok/s or more22 tok/sPass
First replyunder 3 s, 2K prompt4.8 sCheck

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
Hy3
Parameters
295B, 21B active
Precision
2-bit (Q2_K)
Context
32K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
DGX Spark 128 GB
GPU can use
120 GB of 128 GB

Worth trying

  • 8-bit cache: Room for 64K context

Running cost

At home
$3.15/mo
144 W at the wall
Renting instead
$287/mo
4× A40 at $2.36/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.

How to run it

llama.cpp
llama-server -hf bartowski/Hy3-GGUF
vLLM
vllm serve tencent/Hy3
MLX (Mac)
mlx_lm.chat --model mlx-community/Hy3-OptiQ-2bit

Hy3 on similar machines

All machines: Hy3 VRAM requirements. Everything that runs here: what can the DGX Spark run?

Other models that run on DGX Spark

Questions

Can Hy3 run on DGX Spark?

Hy3 at 2-bit runs on the DGX Spark 128 GB with 32K of context at about 22 tokens a second.

Report VR-0OUJOGV, calculated Nov 22, 2026 from published specs, not benchmarked.