Can the NVIDIA A2 run ERNIE 4.5 21B-A3B Thinking?

83 pts

Yes, with little memory to spare

Tight fit. At Q4_K_M it needs about 15.2 GB of 16.0 GB usable, with an estimated speed of 40.9 tok/s.

Best quant: Q4_K_M
Generation speed: 40.9 tok/s
Prompt processing: 1,080 tok/s
First token (1K prompt): 0.9 s
Largest safe context: 21K

Download ERNIE 4.5 21B-A3B Thinking

Recommended for the NVIDIA A2: Q4_K_M. Every other quantization has its own download link in the table below.

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

Download the Q4_K_M file (opens in a new tab)

huggingface.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF (opens in a new tab)

ollama run hf.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M

Original weights · Hugging Face

Full-precision files from Baidu, for fine-tuning, vLLM, Transformers.

huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking (opens in a new tab)

No sign-in needed.

Official repo · GitHub

Reference code, inference scripts and docs from Baidu. The weights themselves download from Hugging Face.

github.com/PaddlePaddle/ERNIE (opens in a new tab)

git clone https://github.com/PaddlePaddle/ERNIE

Links checked 5 days ago against Hugging Face, Ollama and GitHub.

How the A grade is scored

Every result gets up to 100 points from four parts. The points decide the letter grade.

Model qualityEditorial quality tier of the model, scaled by the quantization or precision used.	29.9 / 45
Memory fitFull marks when everything sits in GPU or unified memory; fewer when part spills to system RAM.	24.0 / 25
SpeedGeneration speed against a 40 tokens/s target for LLMs, or render time against a target for image and video.	15.0 / 15
Memory useRewards using the memory you have, so a bigger, better variant beats a tiny one that leaves most memory idle.	15.0 / 15
Total	83.9 / 100	83 ptsGreat (A) · 83 of 100 points: Model quality 29.9/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15

Grade scale: S 85+ · A 72+ · B 58+ · C 42+ · D 25+ · F under 25 points.

Every quantization on the NVIDIA A2

Quant	Memory needed	Available	Speed	Fit	Grade · points	Download
`Q2_K`	10.6 GB	16.0 GB	67.2 tok/s	Runs well	77 ptsGreat (A) · 77 of 100 points: Model quality 22.1/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15	GGUF (opens in a new tab)
`Q3_K_M`	12.6 GB	16.0 GB	52.9 tok/s	Runs well	81 ptsGreat (A) · 81 of 100 points: Model quality 26.5/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15	GGUF (opens in a new tab)
`Q4_K_M`	15.2 GB	16.0 GB	40.9 tok/s	Tight fit	83 ptsGreat (A) · 83 of 100 points: Model quality 29.9/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	GGUF (opens in a new tab)
`Q5_K_M`	17.4 GB	16.0 GB	28.7 tok/s	Runs with offload	68 ptsGood (B) · 68 of 100 points: Model quality 30.6/45, Memory fit 10.0/25, Speed 12.7/15, Memory use 15.0/15	GGUF (opens in a new tab)
`Q6_K`	19.8 GB	16.0 GB	20.0 tok/s	Runs with offload	66 ptsGood (B) · 66 of 100 points: Model quality 31.2/45, Memory fit 10.0/25, Speed 10.6/15, Memory use 15.0/15	GGUF (opens in a new tab)
`Q8_0`	25.0 GB	16.0 GB	11.7 tok/s	Heavy offload	58 ptsGood (B) · 58 of 100 points: Model quality 31.5/45, Memory fit 4.0/25, Speed 8.1/15, Memory use 15.0/15	GGUF (opens in a new tab)
`F16`	45.5 GB	16.0 GB	—	Won't run	9 ptsWon't run (F) · 9 of 100 points: Model quality 9.5/45, Memory fit 0.0/25, Speed 0.0/15, Memory use 0.0/15	GGUF (opens in a new tab) Original weights (opens in a new tab)

Memory figures assume 8K of context. Longer conversations or documents grow the KV cache.

Other hardware that runs it

86 ptsExcellent (S) · 86 of 100 points: Model quality 31.5/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 5090 Q8_0 · 187 tok/s
86 ptsExcellent (S) · 86 of 100 points: Model quality 31.2/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 AMD Radeon RX 7900 XTX Q6_K · 117 tok/s
86 ptsExcellent (S) · 86 of 100 points: Model quality 31.2/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 3090 Q6_K · 133 tok/s
86 ptsExcellent (S) · 86 of 100 points: Model quality 31.2/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15 NVIDIA GeForce RTX 4090 Q6_K · 143 tok/s
84 ptsGreat (A) · 84 of 100 points: Model quality 31.5/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 13.5/15 Apple M4 Max (64 GB) Q8_0 · 49.0 tok/s
83 ptsGreat (A) · 83 of 100 points: Model quality 29.9/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15 AMD Radeon RX 9070 XT Q4_K_M · 113 tok/s

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.