Can Phi 3.5 vision instruct run on the RTX A6000?
Yes. At 8K context the best fit is F16, using about 11.2 GB of the 45.6 GB usable on the RTX A6000. Estimated speed ceiling: ~93 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) | 3.4 GB Fits | 4.9 GB Fits | 13.9 GB Fits | 49.9 GB No | 494 |
| 3-bit (Q3_K_M) | 3.8 GB Fits | 5.3 GB Fits | 14.3 GB Fits | 50.3 GB No | 379 |
| 4-bit (Q4_K_M) | 4.3 GB Fits | 5.8 GB Fits | 14.8 GB Fits | 50.8 GB No | 303 |
| 5-bit (Q5_K_M) | 4.7 GB Fits | 6.2 GB Fits | 15.2 GB Fits | 51.2 GB No | 260 |
| 6-bit (Q6_K) | 5.1 GB Fits | 6.6 GB Fits | 15.6 GB Fits | 51.6 GB No | 226 |
| 8-bit (Q8_0) | 6.1 GB Fits | 7.6 GB Fits | 16.6 GB Fits | 52.6 GB No | 174 |
| 16-bit (F16/BF16) | 9.7 GB Fits | 11.2 GB Fits | 20.2 GB Fits | 56.2 GB No | 93 |
• real GGUF file size. Speed is a bandwidth ceiling (768 GB/s), not a benchmark. Methodology.
More
Phi 3.5 vision instruct on every GPU and Mac · Every model that runs on the RTX A6000 · Other models that fit 48 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (microsoft/Phi-3.5-vision-instruct), NVIDIA specs. Data as of 2026-11-17.