RunsOnMy ModelsGPUs & MacsBy VRAMAPI pricesVRAM calculatorNew

Can I run ERNIE 4.5 21B A3B Thinking locally?

ERNIE 4.5 21B A3B Thinking by baidu has 22B parameters. At 4-bit with an 8K context it needs about 13.5 GB of GPU memory. It fits on 59 of 65 GPUs and Macs we track. See all models that fit in 16 GB of VRAM.

13.5 GBQ4_K_M · 8K context
RTX 4060 Ti 16GBsmallest consumer GPU that fits
M2 24GBsmallest Mac that fits
128Kmax context
GQAattention type

Memory needed by quant and context length

GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use real file sizes from unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF where available (marked •), otherwise parameter count × bits per weight. How we calculate.

QuantQuality4K ctx8K ctx32K ctx128K ctx
2-bit (Q2_K) •rough8.4 GB8.6 GB9.9 GB15.2 GB
3-bit (Q3_K_M) •usable10.6 GB10.8 GB12.1 GB17.4 GB
4-bit (Q4_K_M) •good, most popular13.2 GB13.5 GB14.8 GB20.0 GB
5-bit (Q5_K_M) •very good15.3 GB15.5 GB16.9 GB22.1 GB
6-bit (Q6_K) •near lossless17.6 GB17.8 GB19.1 GB24.3 GB
8-bit (Q8_0) •effectively lossless22.4 GB22.6 GB24.0 GB29.2 GB
16-bit (F16/BF16) •original weights41.5 GB41.7 GB43.0 GB48.3 GB

Which GPUs and Macs run ERNIE 4.5 21B A3B Thinking?

Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).

HardwareMemoryFits?Best quantNeedsMax tok/s (est.)
RTX 3060 12GB12 GBFitsQ2_K8.6 GB44details
RTX 4060 Ti 16GB16 GBFitsQ4_K_M13.5 GB22details
RTX 309024 GBFitsQ6_K17.8 GB52details
RTX 409024 GBFitsQ6_K17.8 GB56details
RTX 509032 GBFitsQ8_022.6 GB77details
RTX 5070 Ti16 GBFitsQ4_K_M13.5 GB67details
RX 7900 XTX24 GBFitsQ6_K17.8 GB53details
Mac M4 Pro 48GB48 GBFitsQ8_022.6 GB12details
Mac M4 Max 128GB128 GBFitsF1641.7 GB13details
Mac M3 Ultra 512GB512 GBFitsF1641.7 GB19details
DGX Spark128 GBFitsF1641.7 GB6.3details
RTX 407012 GBFitsQ2_K8.6 GB62details
RTX 5060 Ti 16GB16 GBFitsQ4_K_M13.5 GB34details
RTX 40608 GBNo———details
Arc B58012 GBFitsQ2_K8.6 GB56details
RTX 3060 Ti8 GBNo———details
RTX 30708 GBNo———details
RTX 3080 10GB10 GBTightQ2_K8.6 GB93details
RTX 3080 12GB12 GBFitsQ2_K8.6 GB112details
RTX 3090 Ti24 GBFitsQ6_K17.8 GB56details
RTX 4060 Ti 8GB8 GBNo———details
RTX 4070 Super12 GBFitsQ2_K8.6 GB62details
RTX 4070 Ti12 GBFitsQ2_K8.6 GB62details
RTX 4070 Ti Super16 GBFitsQ4_K_M13.5 GB50details
RTX 408016 GBFitsQ4_K_M13.5 GB54details
RTX 4080 Super16 GBFitsQ4_K_M13.5 GB55details
RTX 50608 GBNo———details
RTX 507012 GBFitsQ2_K8.6 GB82details
RTX 508016 GBFitsQ4_K_M13.5 GB72details
2x RTX 30902×24 GBFitsQ8_022.6 GB40details
2x RTX 40902×24 GBFitsQ8_022.6 GB43details
2x RTX 50902×32 GBFitsF1641.7 GB41details
RTX A600048 GBFitsQ8_022.6 GB33details
RTX 6000 Ada48 GBFitsQ8_022.6 GB41details
RTX PRO 6000 Blackwell96 GBFitsF1641.7 GB41details
L40S48 GBFitsQ8_022.6 GB37details
A100 40GB40 GBFitsQ8_022.6 GB67details
A100 80GB80 GBFitsF1641.7 GB47details
H100 PCIe80 GBFitsF1641.7 GB46details
H100 SXM80 GBFitsF1641.7 GB77details
H200141 GBFitsF1641.7 GB110details
RX 76008 GBNo———details
RX 7700 XT12 GBFitsQ2_K8.6 GB53details
RX 7800 XT16 GBFitsQ4_K_M13.5 GB47details
RX 7900 GRE16 GBFitsQ4_K_M13.5 GB43details
RX 7900 XT20 GBFitsQ5_K_M15.5 GB51details
RX 907016 GBFitsQ4_K_M13.5 GB48details
RX 9070 XT16 GBFitsQ4_K_M13.5 GB48details
Ryzen AI Max+ 395 128GB128 GBFitsF1641.7 GB5.9details
Arc A770 16GB16 GBFitsQ4_K_M13.5 GB42details
Mac M1 16GB16 GBFitsQ2_K8.6 GB8.3details
Mac M2 24GB24 GBFitsQ4_K_M13.5 GB7.5details
Mac M3 24GB24 GBFitsQ4_K_M13.5 GB7.5details
Mac M4 16GB16 GBFitsQ2_K8.6 GB15details
Mac M4 32GB32 GBFitsQ6_K17.8 GB6.7details
Mac M1 Pro 32GB32 GBFitsQ6_K17.8 GB11details
Mac M3 Pro 36GB36 GBFitsQ6_K17.8 GB8.3details
Mac M4 Pro 24GB24 GBFitsQ4_K_M13.5 GB20details
Mac M4 Pro 64GB64 GBFitsF1641.7 GB6.3details
Mac M1 Max 64GB64 GBFitsF1641.7 GB9.2details
Mac M2 Max 96GB96 GBFitsF1641.7 GB9.2details
Mac M3 Max 128GB128 GBFitsF1641.7 GB9.2details
Mac M4 Max 64GB64 GBFitsF1641.7 GB13details
Mac M2 Ultra 192GB192 GBFitsF1641.7 GB18details
Mac M3 Ultra 256GB256 GBFitsF1641.7 GB19details

Questions

How much VRAM does ERNIE 4.5 21B A3B Thinking need?

About 13.5 GB at 4-bit (Q4_K_M) with an 8K context: 12.4 GB of weights (real GGUF file size), 0.4 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 22.6 GB.

What is the smallest GPU that runs ERNIE 4.5 21B A3B Thinking?

At Q4_K_M and 8K context, the smallest single consumer GPU in our table that fits it is the RTX 4060 Ti 16GB (16 GB).

Can I run ERNIE 4.5 21B A3B Thinking on a Mac?

Yes. At Q4_K_M with 8K context it fits on a Mac M2 24GB or larger, using the default macOS GPU memory limit.

Similar models

Same family or a similar parameter count, with the VRAM each needs at Q4_K_M and 8K context.

ERNIE 4.5 21B A3B PT

22B · 13.5 GB at 4-bit

Qianfan OCR

4.7B · 4.4 GB at 4-bit

Unlimited OCR 2

3.3B · 3.0 GB at 4-bit

gpt oss safeguard 20b

22B · 11.8 GB at 4-bit

gpt oss 20b

21B · 11.8 GB at 4-bit

LFM2.5 24B A2B

24B · 14.3 GB at 4-bit

Model facts

Hugging Face repobaidu/ERNIE-4.5-21B-A3B-Thinking
Parameters21,825,437,888 (22B)
Layers / hidden size28 / 2560
AttentionGQA; 20 heads, 4 KV heads, head dim 128
KV cache per token (FP16)56.0 KiB
Max context131,072 tokens
Licenseapache-2.0
Released on HFSep 8, 2025
Downloads (30d) / likes85,588 / 796

Sources: Hugging Face model API · config.json (baidu/ERNIE-4.5-21B-A3B-Thinking) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.