Compute Engine
Rent Google Cloud GPU and TPU virtual machines and run any open model yourself.
Overview
Compute Engine is Google Cloud's virtual machine service. Its accelerator-optimised machines, from a single NVIDIA T4 up to eight NVIDIA B200s, plus Google's own TPUs, let you run any open-weight model with full control over the software. GPUs are billed per hour on top of the VM, with discounts for committed use and Spot capacity.
Specifications
| Hosting type | GPU cloud |
|---|---|
| Model families | gpt-oss, Qwen, DeepSeek, Llama, Gemma, Mistral |
| LLM models | gpt-oss-120b, Qwen2.5 72B Instruct, DeepSeek-V3, Llama 3.3 70B Instruct, Llama 3.1 8B Instruct, gpt-oss-20b, Gemma 3 27B, Gemma 3 12B, Gemma 3 4B, Gemma 2 27B, Gemma 2 9B, Qwen3 30B-A3B, Qwen3 32B, Qwen3 235B-A22B, Qwen2.5-Coder 32B Instruct, DeepSeek-V3.1, DeepSeek-R1, Codestral 22B, Mistral Small 3.1, Mistral NeMo |
| Pricing range | From US$0.35 |
| Pricing model | Per GPU-hour, Reserved or committed capacity |
| Free tier available | No |
| Data center locations | United States, Canada, Brazil, Chile, Mexico, Belgium, Netherlands, Germany, Finland, United Kingdom, France, Italy, Spain, Switzerland, Poland, Sweden, Israel, Qatar, Saudi Arabia, South Africa, India, Japan, South Korea, Taiwan, Hong Kong SAR China, Singapore, Indonesia, Australia |
| Hardware / GPU options | G2 (NVIDIA L4), A2 (A100), A3 (H100 / H200), A4 (B200) and A4X (GB200) machine types, N1 VMs with NVIDIA T4, plus Google Cloud TPUs. |
| VRAM & memory | 16 GB – 180 GB |
| Maximum model size | 671B |
| OpenAI-compatible API | No |