Can Phi 4 mini instruct run on the L40S?
Yes. At 8K context the best fit is F16, using about 8.9 GB of the 45.6 GB usable on the L40S. Estimated speed ceiling: ~112 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | 2.8 GB Fits | 3.3 GB Fits | 6.3 GB Fits | 18.3 GB Fits | 513 |
| 3-bit (Q3_K_M) • | 3.3 GB Fits | 3.8 GB Fits | 6.8 GB Fits | 18.8 GB Fits | 408 |
| 4-bit (Q4_K_M) • | 3.6 GB Fits | 4.1 GB Fits | 7.1 GB Fits | 19.1 GB Fits | 347 |
| 5-bit (Q5_K_M) • | 3.9 GB Fits | 4.4 GB Fits | 7.4 GB Fits | 19.4 GB Fits | 303 |
| 6-bit (Q6_K) • | 4.2 GB Fits | 4.7 GB Fits | 7.7 GB Fits | 19.7 GB Fits | 274 |
| 8-bit (Q8_0) • | 5.1 GB Fits | 5.6 GB Fits | 8.6 GB Fits | 20.6 GB Fits | 212 |
| 16-bit (F16/BF16) • | 8.4 GB Fits | 8.9 GB Fits | 11.9 GB Fits | 23.9 GB Fits | 112 |
• real GGUF file size from unsloth/Phi-4-mini-instruct-GGUF. Speed is a bandwidth ceiling (864 GB/s), not a benchmark. Methodology.
More
Phi 4 mini instruct on every GPU and Mac · Every model that runs on the L40S · Other models that fit 48 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (microsoft/Phi-4-mini-instruct), NVIDIA specs. Data as of 2026-11-17.