What a smaller model saves
A smaller model fits on a cheaper GPU. In our first test that means up to 35% potential savings on the monthly GPU bill, and we project more for much larger models.
The same model on a cheaper GPU
With its image module, Google's 4-bit Gemma 4 26B is too big for the 16 GB GPU, so it needs the 24 GB one. Uluka Dango fits the 16 GB GPU with images.
Each server runs one copy of the model. The costs and the saving below scale with the number of servers.
Google's 4-bit version
15.6 GB
- GPU it needs
- 24 GB (L4)
- GPU price
- $0.80 an hour
$588 a month
$0.80 an hour × 730 hours
Uluka Dango
8.7 GB
- GPU it needs
- 16 GB (T4)
- GPU price
- $0.53 an hour
$384 a month
$0.53 an hour × 730 hours
Potential saving
$20435% potential saving
a month, or $2,442 a year, against Google's 4-bit setup
GPU prices are Amazon Web Services on-demand, US East, September 2026, for 730 hours a month. This is hourly price only. The two GPUs run at different speeds, so cost per request is not the same as cost per hour.
Why the setups differ
A model has to fit inside its GPU, so each version of Gemma 4 26B needs a different one.
| GPU | Original 51.6 GB | Google 4-bit 15.6 GB | Uluka Dango 8.7 GB |
|---|---|---|---|
| 12 GBRTX 3060, RTX 4070 SUPER | ×Does not fit | ×Does not fit | ✓Fits |
| 16 GBT4 | ×Does not fit | ×Does not fit | ✓Fits |
| 24 GBL4, RTX 4090 | ×Does not fit | ✓Fits | ✓Fits |
| 32 GBRTX 5090 | ×Does not fit | ✓Fits | ✓Fits |
| 48 GBL40S | ×Does not fit | ✓Fits | ✓Fits |
| 80 GBA100, H100 | ✓Fits | ✓Fits | ✓Fits |
Sizes include each model's image module. A model counts as fitting if 1.5 GB is left free for it to work.
A much larger model
At 26B the saving is modest: community builds of 11 to 14 GB already fit on the 16 GB GPU. The saving should grow with model size. Inkling-Small, from Thinking Machines, is a 266B model, and its official builds need two to four 96 GB GPUs. A version that fits on one should cut GPU cost by more than half. That is the target for Uluka 266B.
Projected cost per server, per month
| Inkling-Small version | GPUs needed (96 GB each) | Cost per month* | Uluka 266B would save per month* |
|---|---|---|---|
| Uluka 266Bour target, about 80 GB | 1 | $2,453 projected* | – |
| Q4UD-Q4_K_XL, Unsloth, 163 GB | 2 | $6,052 | $3,599 (59%) |
| NVFP4Thinking Machines, 171 GB | 2 | $6,052 | $3,599 (59%) |
| FP8Red Hat, about 270 GB | 4 | $12,096 | $9,643 (80%) |
*Projected: Uluka 266B is not built yet. Prices are Amazon Web Services g7e instances, on-demand, US East, October 2026: $3.36 an hour for one GPU, $8.29 for two, $16.57 for four. Versions that need the same number of GPUs cost the same. Sizes are the published Hugging Face builds. Community 2-bit builds (about 88 GB) also fit one GPU; our aim is 4-bit quality at a smaller size. Speed is not compared.