ERNIE 4.5 21B-A3B Thinking on Apple M2 Ultra (64 GB)

Can the Apple M2 Ultra (64 GB) run ERNIE 4.5 21B-A3B Thinking?

Yes, it runs great

Runs great. At Q8_0 it needs about 28.0 GB of 48.0 GB usable, with an estimated speed of 92.8 tok/s.

Best quant: Q8_0
Generation speed: 92.8 tok/s
Prompt processing: 5,621 tok/s
First token (1K prompt): 0.3 s
Largest safe context: 128K

Best app to run it on the Apple M2 Ultra (64 GB)

On macOS

LM Studio

Desktop app · runs on llama.cpp, MLX

Point-and-click: search, download and chat in one window.

Open ERNIE 4.5 21B-A3B Thinking in LM Studio

llama.cpp Usually the first to support new model architectures.
Ollama Runs the Hugging Face GGUF with ollama run hf.co/…

How we choose: the files published for this model, your operating system and GPU, whether it fits fully in memory, and published speed tests (MLX vs llama.cpp on Apple silicon, vLLM vs llama.cpp vs Ollama on NVIDIA). Details.

Download ERNIE 4.5 21B-A3B Thinking

Recommended for the Apple M2 Ultra (64 GB): Q8_0. 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 Q8_0 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:Q8_0

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 3 days ago against Hugging Face, Ollama and GitHub.

How it’s rated

Quality 70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 84/100 (quality, fit, speed and memory use) sets the ranking.: How capable the model itself is at Q8_0, from its editorial quality tier scaled by the quantization or precision. It’s the same on any hardware that runs this variant. Scale: A+ 85+ · A 70+ · B 55+ · C 40+ · D under 40.
Runs Runs great: How well it runs on the Apple M2 Ultra (64 GB): speed, capped by how well it fits in memory. Runs great: 30+ tokens/s (or near the image, video or decision speed target) · Runs well: 10+ tokens/s · Slow: 3+ tokens/s · Very slow: under 3 · Won’t run: doesn’t fit even with offload. A tight fit or RAM offload caps it at Runs well; heavy offload caps it at Slow.

Match score

To order lists, each result also gets a match score out of 100 from four parts. It favors the most capable model that still runs well.

Model qualityEditorial quality tier of the model, scaled by the quantization or precision used.	31.5 / 45
Memory fitFull marks when the model fits in the share of unified memory the GPU can use; fewer when it goes past that limit.	25.0 / 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.	13.5 / 15
Match score	85.0 / 100

The match score orders results; it is not shown as a grade. Quality and Runs are the two ratings in the tables.

Every quantization on the Apple M2 Ultra (64 GB)

Quant	Memory needed	Available	Speed	Runs	Quality	Download
`Q2_K`	13.6 GB	48.0 GB	189 tok/s	Runs great	49Fair quality (C) · 49 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 68/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q3_K_M`	15.6 GB	48.0 GB	167 tok/s	Runs great	58Good quality (B) · 58 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 73/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q4_K_M`	18.2 GB	48.0 GB	145 tok/s	Runs great	66Good quality (B) · 66 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 78/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q5_K_M`	20.4 GB	48.0 GB	128 tok/s	Runs great	67Good quality (B) · 67 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q6_K`	22.8 GB	48.0 GB	115 tok/s	Runs great	69Good quality (B) · 69 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`Q8_0`	28.0 GB	48.0 GB	92.8 tok/s	Runs great	70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 84/100 (quality, fit, speed and memory use) sets the ranking.	GGUF (opens in a new tab)
`F16`	48.6 GB	48.0 GB	43.6 tok/s	Runs wellover GPU memory limit	70Great quality (A) · 70 of 100: how capable the model is at F16, whatever it runs on. Match score 71/100 (quality, fit, speed and memory use) sets the ranking.	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

70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. Apple M4 Pro (48 GB) Q8_0 · 41.3 tok/s · Runs great
70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA GeForce RTX 5090 Q8_0 · 154 tok/s · Runs great
69Good quality (B) · 69 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. AMD Radeon RX 7900 XTX Q6_K · 127 tok/s · Runs great
69Good quality (B) · 69 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA GeForce RTX 3090 Q6_K · 136 tok/s · Runs great
69Good quality (B) · 69 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 86/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA GeForce RTX 4090 Q6_K · 141 tok/s · Runs great
70Great quality (A) · 70 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 84/100 (quality, fit, speed and memory use) sets the ranking. Apple M1 Max (64 GB) Q8_0 · 56.4 tok/s · Runs great

Quick answers

How much memory does ERNIE 4.5 21B-A3B Thinking need on the M2 Ultra 64GB?

At Q8_0 it needs about 28.0 GB. The M2 Ultra 64GB has 48.0 GB usable.

How fast is ERNIE 4.5 21B-A3B Thinking on the M2 Ultra 64GB?

About 92.8 tok/s at Q8_0, an estimate from the hardware’s memory bandwidth and the size of the model.

Which quantization of ERNIE 4.5 21B-A3B Thinking should I use on the M2 Ultra 64GB?

Q8_0. Of the versions that fit, it is the one that runs best; when two run equally well, the check prefers the higher-rated one.

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.