AI Models We Recommend
These are three real, current, open-source frontier models — all 100B+ parameters, all under an open license that lets you self-host and use them commercially.
How we calculate memory needs: parameters (in billions) × 1 GB × 1.2 (a 20% safety buffer
for the "scratch space" the model needs while actually answering a question — things like conversation
history and temporary calculations). This approximates running the model at roughly 8-bit precision.
Z.ai · MIT (fully open)
GLM 5.2 Max
Total parameters
753 B
Active parameters per response
40 B
Mixture-of-experts model — 753B total parameters, but only about 40B are "active" per response, which is why it can be faster than its total size suggests.
Memory math
753 × 1 GB × 1.2
Memory needed
903.6 GB
DeepSeek · MIT (fully open)
DeepSeek V4 Pro
Total parameters
1,600 B
Active parameters per response
49 B
The largest of the three — 1.6 trillion total parameters, mixture-of-experts with ~49B active per response.
Memory math
1600 × 1 GB × 1.2
Memory needed
1,920 GB
Moonshot AI · Modified MIT (open, commercial use allowed with attribution)
Kimi K2.7 Code
Total parameters
1,000 B
Active parameters per response
32 B
Specialized for coding and long, multi-step agent tasks. 1 trillion total parameters, ~32B active per response, the leanest "active" footprint of the three.
Memory math
1000 × 1 GB × 1.2
Memory needed
1,200 GB
How Many GPUs Does That Actually Take?
Using the NVIDIA H100 (80 GB of memory each) as a reference point, rounding up to whole GPUs:
| Model | Memory Needed | H100s Required (80GB each) | Fits in Rack Servers (8 GPUs each) |
|---|---|---|---|
| GLM 5.2 Max | 903.6 GB | 12 GPUs | 2 servers (16 GPU capacity) |
| DeepSeek V4 Pro | 1,920 GB | 24 GPUs | 3 servers (24 GPU capacity) |
| Kimi K2.7 Code | 1,200 GB | 15 GPUs | 2 servers (16 GPU capacity) |
These are big numbers on purpose — all three models are frontier-scale. See our Startup Guide for how smaller teams can run a compressed version instead.