How much VRAM does ERNIE 4.5 300B-A47B Instruct need?
ERNIE 4.5 300B-A47B Instruct needs about 178.6 GB of VRAM at Q4_K_M, or 312.5 GB at Q8_0 and 587.5 GB at FP16. No single GPU in our set holds it even at Q4_K_M; it needs multi-GPU or offload. Weights plus runtime overhead. Budget separately for KV cache, which at 128K is a material share.
Weights-plus-runtime footprint across the six most common quantizations. KV cache is context-dependent and comes on top. At 128K it is gigabytes on its own, before the weights; size it precisely in the GPU Memory Calculator.
This is a mixture-of-experts model: all 300B parameters must sit in memory, but only about 47B are active per token — so it loads like a 300B model and runs closer to a 47B one.
Can your GPU run ERNIE 4.5 300B-A47B Instruct?
Pick your card — the answer below is computed by the same engine that produces every figure on this page.
H200 141GB cannot hold ERNIE 4.5 300B-A47B Instruct at Q4_K_M: it needs 178.6 GB against 133.9 GB usable; a heavier quantization, a larger card, or splitting across GPUs are the ways forward; as a mixture-of-experts model all 300B parameters must be resident even though only 47B are active per token.
Weights plus runtime overhead. KV cache is not included — it depends on context length and this model’s published architecture.
| Weights (Q4_K_M) | 169.4 GB |
| Runtime overhead | 9.2 GB |
| Estimated total | 178.6 GB |
| Usable VRAM (95% of 141 GB) | 133.9 GB |
Across every GPU we track
- RTX 3060 12GBNot enough-167.2 GB
- RTX 4060 Ti 16GBNot enough-163.4 GB
- RTX 4070 Ti SUPERNot enough-163.4 GB
- RTX 3090Not enough-155.8 GB
- RTX 4090Not enough-155.8 GB
- RTX 5090Not enough-148.2 GB
- A100 40GBNot enough-140.6 GB
- RTX A6000Not enough-133.0 GB
- L40SNot enough-133.0 GB
- A100 80GBNot enough-102.6 GB
- H100 80GBNot enough-102.6 GB
- H200 141GBNot enough-44.7 GB
Estimates, not guarantees: real usage moves with runtime, driver, batch size and context. Size a specific context window in the GPU Memory Calculator.
How was this number calculated? Every figure here comes from Bitpute’s documented calculation methodology — parameters, precision, quantization, runtime overhead and usable VRAM, each formula written out in full.
Memory by quantization
| Quant | Weights | + runtime overhead |
|---|---|---|
| FP16 | 558.8 GB | 587.5 GB |
| Q8_0 | 296.9 GB | 312.5 GB |
| Q6_K | 229.1 GB | 241.3 GB |
| Q5_K_M | 198.7 GB | 209.4 GB |
| Q4_K_M | 169.4 GB | 178.6 GB |
| Q4_0 | 157.2 GB | 165.8 GB |
Overhead = 0.75 GB + 5% of weights (CUDA context, buffers). Add KV cache on top. Weights dominate at this size, but a 128K window still costs gigabytes.
Which quantization should you actually run?
ERNIE 4.5 300B-A47B Instruct is a multi-GPU or datacenter model: even Q4_K_M needs about 178.6 GB, beyond any single card. Plan for several 80 GB GPUs or a hosted endpoint; FP16 (587.5 GB) is server-only. As a mixture-of-experts model it must hold all 300B in VRAM but activates only about 47B per token, so it runs much faster than its footprint suggests — memory is the limit here, not speed.
Single-GPU fit at Q4_K_M (178.6 GB + KV)
RTX 3060 12GBRTX 4060 Ti 16GBRTX 4070 Ti SUPERRTX 3090RTX 4090RTX 5090A100 40GBRTX A6000L40SA100 80GBH100 80GBH200 141GB
Common questions
How much VRAM does ERNIE 4.5 300B-A47B Instruct need?
ERNIE 4.5 300B-A47B Instruct needs about 178.6 GB of VRAM at Q4_K_M, 312.5 GB at Q8_0, or 587.5 GB at FP16. That is weights plus about 0.75 GB of runtime overhead and 5% of weight size; KV cache is additional and depends on context length.
What GPU can run ERNIE 4.5 300B-A47B Instruct?
No single GPU in our set fits it at Q4_K_M; it needs multiple GPUs or CPU offload. At FP16 no single GPU in our set is sufficient.
Can ERNIE 4.5 300B-A47B Instruct run on 24 GB?
Not at Q4_K_M: it needs about 178.6 GB, more than the ~22.8 GB usable on a 24 GB card. Use a larger card, multiple GPUs, or offload.