Can ERNIE 4.5 21B A3B Thinking run on the RTX 3070?
Not fully in VRAM. Even at 2-bit with 8K context, ERNIE 4.5 21B A3B Thinking needs about 8.6 GB, more than the 7.6 GB usable on the RTX 3070. It can still run with offload (slower): Q4_K_M with part of the model in system RAM, about 3.2–6.2 tokens/sec (estimate). Details and command.
Runs with offload (slower)
Runs with offload 4-bit (Q4_K_M) at 8K context: 14 of 28 layers on the RTX 3070 (about 7.0 GB) and the other 14 in system RAM (about 6.4 GB). Needs a PC with at least 32 GB of system RAM. Estimated speed: 3.2–6.2 tokens/sec (slow but usable), versus a bandwidth ceiling of ~34 tok/s if it fit fully in VRAM.
llama-server -hf unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M -c 8192 -ngl 14
Estimate, not a benchmark. The command downloads the GGUF from Hugging Face. The speed range spans dual-channel DDR4-3200 to DDR5-6000 system RAM; your CPU, RAM speed and llama.cpp version move it. If llama.cpp runs out of GPU memory, lower -ngl by 2-3; if VRAM is left over, raise it for more speed. Recent llama.cpp builds can also pick the split automatically. How we estimate offload.
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 No | 8.6 GB No | 9.9 GB No | 15.2 GB No | — |
| 3-bit (Q3_K_M) • | 10.6 GB No | 10.8 GB No | 12.1 GB No | 17.4 GB No | — |
| 4-bit (Q4_K_M) • | 13.2 GB No | 13.5 GB No | 14.8 GB No | 20.0 GB No | — |
| 5-bit (Q5_K_M) • | 15.3 GB No | 15.5 GB No | 16.9 GB No | 22.1 GB No | — |
| 6-bit (Q6_K) • | 17.6 GB No | 17.8 GB No | 19.1 GB No | 24.3 GB No | — |
| 8-bit (Q8_0) • | 22.4 GB No | 22.6 GB No | 24.0 GB No | 29.2 GB No | — |
| 16-bit (F16/BF16) • | 41.5 GB No | 41.7 GB No | 43.0 GB No | 48.3 GB No | — |
• real GGUF file size from unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF. Speed is a bandwidth ceiling (448 GB/s), not a benchmark. Methodology.
Hardware that does fit it
RTX 3080 10GBRTX 3060 12GBRTX 3080 12GBRTX 4070RTX 4070 SuperRTX 4070 Ti
What fits instead
- Cheapest listed GPU that runs ERNIE 4.5 21B A3B Thinking: Arc A770 16GB (16 GB, launch MSRP $349) at Q4_K_M, fully in VRAM at 8K context.
- Largest popular model that fits the RTX 3070: Qwen3.5 9B (9.7B) at Q4_K_M, about 6.4 GB at 8K context.
"Runs" and "fits" mean the whole model in GPU memory at 8K context (4-bit or better for the model pick). Prices are US launch MSRPs, not today's street prices. "Popular" means the 60 most-downloaded models we track.
More
ERNIE 4.5 21B A3B Thinking on every GPU and Mac · Every model that runs on the RTX 3070 · Other models that fit 8 GB VRAM · Try other settings in the calculator
Guides: Best GPU for local LLMs by budget · Can I run gpt-oss, Llama 70B or Qwen3? · Mac vs PC for local LLMs
Data: Hugging Face Hub (baidu/ERNIE-4.5-21B-A3B-Thinking), NVIDIA specs. Model data last checked 2026-11-17; refreshed hourly.