Qwen3.5 27B: VRAM needs and speed by GPU
At Q4_K_M, Qwen3.5 27B needs about 18.2 GB of VRAM. We have 4 measured speed results for it.
26.9B parameters · license: apache-2.0 · GGUF files on Hugging Face
Measured speeds
| Chip | Generation tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Intel Arc Pro B70 | 20.4 | 718 | Q4_K_M | SYCL FA on | tg128 | PMZFX intel-arc-pro-b70-benchmarks (MIT) details |
| Intel Arc Pro B70 | 15.3 | 776 | Q8_0 | SYCL FA on | tg128 | PMZFX intel-arc-pro-b70-benchmarks (MIT) details |
| Intel Arc Pro B70 | 15.1 | 786 | Q6_K | SYCL FA on | tg128 | PMZFX intel-arc-pro-b70-benchmarks (MIT) details |
| Intel Arc Pro B70 | 6.3 | 361 | IQ4_NL | SYCL FA on | tg128 | PMZFX intel-arc-pro-b70-benchmarks (MIT) details |
Memory needed per quantization
| Quant | File size | Needs (estimate) | Smallest card it fits |
|---|---|---|---|
| UD-IQ2_XXS | 8.6 GB | 10.0 GB | 12 GB cards |
| UD-IQ2_M | 10.2 GB | 11.7 GB | 16 GB cards |
| UD-Q2_K_XL | 11.2 GB | 12.7 GB | 16 GB cards |
| UD-IQ3_XXS | 11.5 GB | 13.0 GB | 16 GB cards |
| Q3_K_S | 12.3 GB | 13.8 GB | 16 GB cards |
| Q3_K_M | 13.5 GB | 15.0 GB | 16 GB cards |
| UD-Q3_K_XL | 14.4 GB | 15.9 GB | 24 GB cards |
| IQ4_XS | 15.0 GB | 16.4 GB | 24 GB cards |
| IQ4_NL | 15.7 GB | 17.2 GB | 24 GB cards |
| Q4_0 | 15.7 GB | 17.2 GB | 24 GB cards |
| Q4_K_S | 15.8 GB | 17.2 GB | 24 GB cards |
| Q4_K_M | 16.7 GB | 18.2 GB | 24 GB cards |
| Q4_1 | 17.2 GB | 18.7 GB | 24 GB cards |
| UD-Q4_K_XL | 17.6 GB | 19.1 GB | 24 GB cards |
| Q5_K_S | 18.9 GB | 20.4 GB | 24 GB cards |
| Q5_K_M | 19.6 GB | 21.1 GB | 24 GB cards |
| UD-Q5_K_XL | 20.2 GB | 21.6 GB | 24 GB cards |
| Q6_K | 22.5 GB | 23.9 GB | 32 GB cards |
| UD-Q6_K_XL | 25.7 GB | 27.1 GB | 32 GB cards |
| Q8_0 | 28.6 GB | 30.1 GB | 32 GB cards |
| UD-Q8_K_XL | 35.5 GB | 37.0 GB | 48 GB cards |
| BF16 | 53.8 GB | 55.3 GB | 96 GB cards |
Expected speed on popular hardware
| Hardware | Best quant that fits | Expected tok/s |
|---|---|---|
| NVIDIA RTX 3060 | UD-IQ2_XXS | 35.9 estimate |
| NVIDIA RTX 4060 Ti | Does not fit | |
| NVIDIA RTX 3090 | Q5_K_M | 33.5 estimate |
| NVIDIA RTX 4090 | Q5_K_M | 44.4 estimate |
| NVIDIA RTX 5090 | Q8_0 | 40.2 estimate |
| AMD Radeon RX 7900 XTX | Q5_K_M | 37.2 estimate |
| Apple M4 Pro (20-core GPU) (64 GB) | Q8_0 | 8.1 estimate |
| Apple M3 Max (40-core GPU) (64 GB) | Q8_0 | 11.8 estimate |
| AMD Ryzen AI Max+ 395 (Radeon 8060S) | Q8_0 | 7.3 estimate |
Too big for your card?
Run Qwen3.5 27B 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.