Can ERNIE 4.5 21B A3B Thinking run on the RTX 6000 Ada?
Yes. At 8K context the best fit is Q8_0, using about 22.6 GB of the 45.6 GB usable on the RTX 6000 Ada. Estimated speed ceiling: ~41 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) • | 8.4 GB Fits | 8.6 GB Fits | 9.9 GB Fits | 15.2 GB Fits | 118 |
| 3-bit (Q3_K_M) • | 10.6 GB Fits | 10.8 GB Fits | 12.1 GB Fits | 17.4 GB Fits | 91 |
| 4-bit (Q4_K_M) • | 13.2 GB Fits | 13.5 GB Fits | 14.8 GB Fits | 20.0 GB Fits | 72 |
| 5-bit (Q5_K_M) • | 15.3 GB Fits | 15.5 GB Fits | 16.9 GB Fits | 22.1 GB Fits | 62 |
| 6-bit (Q6_K) • | 17.6 GB Fits | 17.8 GB Fits | 19.1 GB Fits | 24.3 GB Fits | 53 |
| 8-bit (Q8_0) • | 22.4 GB Fits | 22.6 GB Fits | 24.0 GB Fits | 29.2 GB Fits | 41 |
| 16-bit (F16/BF16) • | 41.5 GB Tight | 41.7 GB Tight | 43.0 GB Tight | 48.3 GB No | 22 |
• real GGUF file size from unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF. Speed is a bandwidth ceiling (960 GB/s), not a benchmark. Methodology.
More
ERNIE 4.5 21B A3B Thinking on every GPU and Mac · Every model that runs on the RTX 6000 Ada · Other models that fit 48 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (baidu/ERNIE-4.5-21B-A3B-Thinking), NVIDIA specs. Data as of 2026-11-17.