Hardware matrix
Compare models across hardware
Which local models run on which machines. Y runs comfortably, ≈ is a tight fit, N needs more memory. Computed from the same validated memory math as every can-i-run page.
Choosing between your hardware and a hosted API? Compare local models with Astra, Claude and Gemini using Arena preference ratings, memory fit and monthly token costs.
| Model | Nvidia GeForce RTX 2060 6 GB | Generic Android Phone 8 GB | Samsung Galaxy S24 Ultra 12 GB | Nvidia GeForce RTX 4060 Ti 16 GB | Apple M5 16 GB | Nvidia GeForce RTX 4090 24 GB | Apple M4 Pro 48 GB | Apple M3 Ultra 256 GB |
|---|---|---|---|---|---|---|---|---|
| SmolLM2 135M | Y | Y | Y | Y | Y | Y | Y | Y |
| Qwen2.5 1.5B | Y | Y | Y | Y | Y | Y | Y | Y |
| Phi-3.5-mini 3.8B | Y | ≈ | Y | Y | Y | Y | Y | Y |
| RN RNJ-1 8B | N | N | Y | Y | Y | Y | Y | Y |
| Qwen2.5 14B | N | N | N | Y | N | Y | Y | Y |
| Gemma 4 26B-A4B | N | N | N | N | N | Y | Y | Y |
| DeepSeek-R1-Distill-Qwen 32B | N | N | N | N | N | ≈ | Y | Y |
| SE Seed-OSS 36B Instruct | N | N | N | N | N | ≈ | Y | Y |
| DeepSeek-V4-Flash | N | N | N | N | N | N | N | ≈ |
| KI Kimi K3 | N | N | N | N | N | N | N | N |
This is a representative slice. For any exact pairing across all 149 models and 43 devices, open its can-i-run page or use the detector.
FAQ
How do I compare local AI models across different hardware?
Match each model's memory need at Q4_K_M to the device's usable memory. This matrix does it for 10 representative models across 8 devices; Y means it runs comfortably, ≈ means a tight fit, N means not enough memory. For any exact pairing across all 149 models and 43 devices, open its can-i-run page.