Open-Weight Models
10 profiles in this group. Use this hub page to compare practical VRAM floor, expected throughput, and best local-vs-cloud path.
RAM planning is included because Ollama may spill layers to CPU when VRAM is tight, especially with long context.
| Model | Data | VRAM min | VRAM optimal | System RAM | RTX 3090 fit | Best local GPU | Cloud fallback | Detail |
|---|---|---|---|---|---|---|---|---|
| GPT-OSS 120B CLOUD | Estimated | 70GB | 78GB | 128GB+ | Cloud / larger GPU | Dual RTX 4090 (model parallel) | A100 80GB | Open |
| GPT-OSS 120B FP16 | Estimated | 80GB | 92GB | 128GB+ | Cloud / larger GPU | Cloud-first (no practical single-GPU local) | H100/H200 class | Open |
| GPT-OSS 120B Q4 | Estimated | 70GB | 80GB | 128GB+ | Cloud / larger GPU | Dual RTX 4090 (model parallel) | A100 80GB | Open |
| GPT-OSS 120B Q5 | Estimated | 80GB | 82GB | 128GB+ | Cloud / larger GPU | Cloud-first (no practical single-GPU local) | H100/H200 class | Open |
| GPT-OSS 120B Q8 | Estimated | 74GB | 84GB | 128GB+ | Cloud / larger GPU | Cloud-first (no practical single-GPU local) | H100/H200 class | Open |
| GPT-OSS 20B CLOUD | Estimated | 20GB | 28GB | 64GB+ | Borderline | RTX 6000 Ada 48GB | A100 80GB | Open |
| GPT-OSS 20B FP16 | Estimated | 30GB | 42GB | 64GB+ | Cloud / larger GPU | RTX 6000 Ada 48GB | A100 80GB | Open |
| GPT-OSS 20B Q4 | Estimated | 16GB | 20GB | 32GB+ | Comfortable | RTX 3090 24GB | A6000 48GB | Open |
| GPT-OSS 20B Q5 | Estimated | 20GB | 22GB | 32GB+ | Comfortable | RTX 3090 24GB | A6000 48GB | Open |
| GPT-OSS 20B Q8 | Estimated | 24GB | 34GB | 64GB+ | Borderline | RTX 6000 Ada 48GB | A100 80GB | Open |
We may earn a commission if you click links on this page.