Models › Phi-4-mini-instruct
Phi-4-mini-instruct: VRAM needs and speed by GPU
At Q4_K_M, Phi-4-mini-instruct needs about 3.4 GB of VRAM. No measured runs yet. Speeds below are estimates. Submit a run.
3.84B parameters · license: mit · 131,072-token context · GGUF published Feb 28, 2025 · GGUF files on Hugging Face
Memory needed per quantization
| Quant | File size | Needs (estimate) | Smallest card it fits |
|---|---|---|---|
| Q2_K | 1.7 GB | 2.6 GB | 8 GB cards |
| Q2_K_L | 1.7 GB | 2.6 GB | 8 GB cards |
| Q3_K_M | 2.1 GB | 3.0 GB | 8 GB cards |
| Q4_K_M | 2.5 GB | 3.4 GB | 8 GB cards |
| Q5_K_M | 2.8 GB | 3.7 GB | 8 GB cards |
| Q6_K | 3.2 GB | 4.1 GB | 8 GB cards |
| Q8_0 | 4.1 GB | 5.0 GB | 8 GB cards |
| BF16 | 7.7 GB | 8.6 GB | 12 GB cards |
Expected speed on popular hardware
| Hardware | Best quant that fits | Expected tok/s |
|---|---|---|
| NVIDIA RTX 3060 | Q8_0 | 75.4 estimate |
| NVIDIA RTX 4060 Ti | Q8_0 | 59.9 estimate |
| NVIDIA RTX 3090 | Q8_0 | 161 estimate |
| NVIDIA RTX 4090 | Q8_0 | 178 estimate |
| NVIDIA RTX 5090 | Q8_0 | 281 estimate |
| AMD Radeon RX 7900 XTX | Q8_0 | 179 estimate |
| Apple M4 Pro (20-core GPU) (64 GB) | Q8_0 | 56.7 estimate |
| Apple M3 Max (40-core GPU) (64 GB) | Q8_0 | 82.8 estimate |
| AMD Ryzen AI Max+ 395 (Radeon 8060S) | Q8_0 | 50.9 estimate |
Too big for your card?
Run Phi-4-mini-instruct on a rented 24-80 GB GPU by the hour.
Rent a bigger GPU on RunPodAffiliate link: if you sign up, RigBench may earn a referral credit. It never affects rankings or numbers.