BaiduERNIEErnie4_5_MoeForCausalLM

ERNIE 4.5 21B A3B Thinking — Hardware Requirements & GPU Compatibility

ChatFunctions

ERNIE 4.5 21B A3B Thinking is a 21.8B-parameter open language model from Baidu in the ERNIE family. It supports a context window of up to 131,072 tokens. At Q4_K_M it needs about 13.51 GB of VRAM — see which GPUs and Macs can run it below.

21.8B params · MoE3.4B active131K contextApache 2.0Sep 202558.6K downloads795 likes10.2K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 8 GB (IQ2_XS · 16K)Best on 24 GB (Q4_K_M · ~213 tok/s · 128K)Macs from 16 GB (Q3_K_L · ~55 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run ERNIE 4.5 21B A3B Thinking

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Your own private copy · Q4_K_M · 24 GB GPU

The easiest way to try this model
Running in a few minutes, easy setup
Chat right here or connect any app

$5+ bonus for new users

Run ERNIE 4.5 21B A3B Thinking

14.3 GB VRAM at 16K context · fits the RTX 4090 (9.7 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve baidu/ERNIE-4.5-21B-A3B-Thinking --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~48 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model qurk41/ERNIE-4.5-21B-A3B-Thinking-mlx-4Bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model qurk41/ERNIE-4.5-21B-A3B-Thinking-mlx-4Bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX 3090 (24 GB)CheapestCheapest9 GB free	~202 tok/s	$0.27	Deploy on Vast.aifrom $0.20/hr (opens in new tab)	24 GB · 9 GB free · ~202 tok/s · $0.27 / 1M tokens
RTX 4000 Ada (20 GB)5 GB free	~96 tok/s	$0.54	Deploy on Vast.aifrom $0.19/hr (opens in new tab)	20 GB · 5 GB free · ~96 tok/s · $0.54 / 1M tokens
TITAN RTX (24 GB)9 GB free	~159 tok/s	$0.38	Deploy on Vast.aifrom $0.22/hr (opens in new tab)	24 GB · 9 GB free · ~159 tok/s · $0.38 / 1M tokens
RTX A5000 (24 GB)9 GB free	~176 tok/s	$0.36	Deploy on Vast.aifrom $0.23/hr (opens in new tab)	24 GB · 9 GB free · ~176 tok/s · $0.36 / 1M tokens
RTX PRO 4500 (32 GB)17 GB free	~196 tok/s	$0.48	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 17 GB free · ~196 tok/s · $0.48 / 1M tokens
RTX 4090 (24 GB)9 GB free	~213 tok/s	$0.44	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 9 GB free · ~213 tok/s · $0.44 / 1M tokens
A40 (48 GB)33 GB free	~163 tok/s	$0.60	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 33 GB free · ~163 tok/s · $0.60 / 1M tokens
L4 (24 GB)9 GB free	~82 tok/s	$1.49	Deploy on RunPodfrom $0.44/hr (opens in new tab)	24 GB · 9 GB free · ~82 tok/s · $1.49 / 1M tokens

Compare GPUs →

Your own private copy of ERNIE 4.5 21B A3B Thinking · Q4_K_M · OpenAI-compatible API

Open the llama.cpp listing on RunPod
Open on RunPodServerless (opens in new tab)
New to RunPod? Sign up and add credit first.
2
Fill in the form
Press Deploy and fill in:
Model
unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF
Quantization
Q4_K_M
Context Size
16384
After pasting the model press Esc — don’t pick a suggestion. Press Next.
Create the endpoint
Keep the default 24 GB and 48 GB, then Create Endpoint. Right away: Manage → Edit endpoint → Max workers 1 → Save (changing settings later restarts the worker).
The first start downloads the model, usually 5–15 min. After that, replies after a pause come in seconds.
Create an API key
RunPod → Settings → API Keys → Create. Pick “Restricted” with access to this endpoint only.

Use it

Copy the endpoint URL from its page on RunPod (any of its links works) and paste it with the key. Both stay in this browser.

Endpoint URL

API key

Chats and keys stay only in this browser.

https://api.runpod.ai/v2/<ENDPOINT_ID>/openai/v1

Put your endpoint ID in place of <ENDPOINT_ID>. The API key is the one from step 2.

from openai import OpenAI
client = OpenAI(base_url="https://api.runpod.ai/v2/<ENDPOINT_ID>/openai/v1", api_key="rpa_...")
r = client.chat.completions.create(
    model="any",
    messages=[{"role": "user", "content": "Hello!"}])
print(r.choices[0].message.content)

Works with any OpenAI client.

https://api.runpod.ai/v2/<ENDPOINT_ID>/openai/v1

Put your endpoint ID in place of <ENDPOINT_ID>. The API key is the one from step 2.

Free on Windows, macOS, Linux, iOS, Android and the web.

Open the web app ↗ (opens in a new tab)Download ↗ (opens in a new tab)

1. Settings → Model Provider → Add → Add Custom Provider
2. API Mode OpenAI API Compatible; paste the API key and the base URL as API Host
3. Model → Fetch, pick the model
4. New chat → pick the provider → start chatting

Free and open source for Windows, macOS and Linux.

Download Jan ↗ (opens in a new tab)

1.Settings → next to Model Providers press +
2.API format OpenAI-compatible; Base URL is the base URL above (with /v1); API key is your key; Create
3.If no model appears, click + in Models and type any
4.New chat → pick the model → start chatting

How Much VRAM Does ERNIE 4.5 21B A3B Thinking Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
6.4 GB	8.2 GB	13.8 GB	7.8 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 128K context · ~328 tok/s
7.0 GB	8.7 GB	14.4 GB	6.3 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 128K context · ~315 tok/s
7.2 GB	9 GB	14.6 GB	6.4 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 128K context · ~309 tok/s
7.8 GB	9.5 GB	15.2 GB	8 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 128K context · ~297 tok/s
8.9 GB	10.6 GB	16.3 GB	9.5 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 128K context · ~276 tok/s
9.2 GB	10.9 GB	16.6 GB	7.4 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 128K context · ~271 tok/s
9.4 GB	11.2 GB	16.8 GB	9.4 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 128K context · ~267 tok/s
9.7 GB	11.5 GB	17.1 GB	9.5 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 128K context · ~262 tok/s
9.7 GB	11.5 GB	17.1 GB	8.2 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 128K context · ~262 tok/s
10.0 GB	11.7 GB	17.4 GB	9.6 GB	3-bit small — 3-bit small quantization	Fits: 128K context · ~258 tok/s
10.2 GB	12 GB	17.6 GB	10.3 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 128K context · ~254 tok/s
11.1 GB	12.8 GB	18.5 GB	10.5 GB	3-bit medium — 3-bit medium quantization	Fits: 128K context · ~242 tok/s
11.3 GB	13.1 GB	18.7 GB	12.5 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 128K context · ~238 tok/s
11.6 GB	13.4 GB	19 GB	10.7 GB	3-bit large — 3-bit large quantization	Fits: 128K context · ~235 tok/s
12.2 GB	13.9 GB	19.6 GB	11.8 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 128K context · ~228 tok/s
12.7 GB	14.5 GB	20.1 GB	13.8 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 128K context · ~222 tok/s
12.7 GB	14.5 GB	20.1 GB	12.5 GB	4-bit small — 4-bit small quantization	Fits: 128K context · ~222 tok/s
12.7 GB	14.5 GB	20.1 GB	12.5 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 128K context · ~222 tok/s
13.5 GB	15.3 GB	20.9 GB	13.3 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 128K context · ~213 tok/s
13.8 GB	15.6 GB	21.2 GB	13.6 GB	4-bit large — 4-bit large quantization	Fits: 128K context · ~210 tok/s
14.1 GB	15.8 GB	21.5 GB	15.1 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 128K context · ~207 tok/s
15.4 GB	17.2 GB	22.8 GB	15.1 GB	5-bit small — 5-bit small quantization	Fits: 128K context · ~194 tok/s
16.0 GB	17.7 GB	23.4 GB	15.6 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 128K context · ~189 tok/s
16.2 GB	18 GB	23.6 GB	15.8 GB	5-bit large — 5-bit large quantization	Fits: 128K context · ~187 tok/s
18.4 GB	20.2 GB	25.8 GB	18 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 64K context · ~171 tok/s
22.2 GB	24 GB	29.6 GB	23.2 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Fits: 32K context · ~148 tok/s
44.1 GB	45.8 GB	51.5 GB	43.7 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 44.9 GB

Selected Q4_K_M

4-bit medium quantization

14.3GBat 16K context

Weights: 13.1 GB
KV cache: 0.9 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 9.7 GB free · ~213 tok/s

≥50 tok/s needs ≥173 GB/s · RTX 5090 ~301, RTX 3090 Ti ~213

Which Hardware Can Run ERNIE 4.5 21B A3B Thinking?

Q4_K_M · 13.5 GB

ERNIE 4.5 21B A3B Thinking (Q4_K_M) requires 13.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 18+ GB is recommended. Using the full 131K context window can add up to 7.4 GB, bringing total usage to 20.9 GB. 26 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 124 devices with unified memory can run ERNIE 4.5 21B A3B Thinking, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), MacBook Pro 14" M3 Pro (18 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060IQ2_XS · 7.8 GB · ~132 tok/s · 16K7.8~13216K8 GB tier →

12 GB · RTX 3060 12GBQ3_K_M · 11.9 GB · ~114 tok/s · 16K11.9~11416K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 14.3 GB · ~115 tok/s · 32K14.3~11532K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 14.3 GB · ~213 tok/s · 128K14.3~213128K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 14.3 GB · ~33 tok/s · 128K14.3~33128KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 14.3 GB · ~91 tok/s · 128K14.3~91128KMacs →

Alternatives to ERNIE 4.5 21B A3B Thinking

ERNIE family · 7 models →

Similar models

5tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
InternVL3 5 GPT OSS 20B A4B Preview HFOpenGVLab	—	13.6 GB	~180	same size
Tri 21B Thinktrillionlabs	—	45.1 GB	—needs > 24 GB	same size
EuroLLM 22B Instruct 2512utter-project	—	17.5 GB	~46	same size
Mistral Small Instruct 2409Mistral AI	—	17.4 GB	~46	same size
GPT Neox 20BEleutherAI	—	13.7 GB	~48	same size

Specifications

Publisher: Baidu
Family: ERNIE
Parameters: 21.8B
Active parameters: 3.4B per token
Architecture: Ernie4_5_MoeForCausalLM
Context Length: 131,072 tokens
Vocabulary Size: 103,424
Release Date: 2025-09-08
License: Apache 2.0

Resources

Frequently Asked Questions

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

ERNIE 4.5 21B A3B Thinking requires 13.5 GB of VRAM at Q4_K_M, or 44.1 GB at BF16. Full 131K context adds up to 7.4 GB (20.9 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 21.8B × 4.8 bits ÷ 8 = 13.1 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 7.8 GB (at full 131K context)

VRAM usage by quantization

Q4_K_M

13.5 GB

Q4_K_M + full context

20.9 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run ERNIE 4.5 21B A3B Thinking?

Yes, at Q8_0 (22.2 GB) or lower. Higher quantizations like BF16 (44.1 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for ERNIE 4.5 21B A3B Thinking?

For ERNIE 4.5 21B A3B Thinking, Q4_K_M (13.5 GB) offers the best balance of quality and VRAM usage. Q4_K_L (13.8 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 6.4 GB.

VRAM requirement by quantization

IQ2_XXS

6.4 GB~53%

IQ3_S

9.7 GB~75%

Q3_K_L

11.6 GB~86%

Q4_K_M ★

13.5 GB~89%

Q5_0

14.1 GB~90%

BF16

44.1 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

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

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run ERNIE 4.5 21B A3B Thinking, the smallest at Q3_K_L. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 13.5 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).

Can I run ERNIE 4.5 21B A3B Thinking locally?

Yes — ERNIE 4.5 21B A3B Thinking can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is ERNIE 4.5 21B A3B Thinking?

At Q4_K_M, ERNIE 4.5 21B A3B Thinking can reach ~166 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~213 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 2.1 GB active ÷ (8000 × 0.65) = 0.399 ms, plus 28 layers × 0.055 ms = 1.540 ms, so 1000 ÷ 1.939 ms = ~516 tok/s

Estimated speed at Q4_K_M (13.5 GB)

NVIDIA B200

~516 tok/s

NVIDIA GeForce RTX 4090

~213 tok/s

NVIDIA B300

~516 tok/s

NVIDIA GH200 Grace Hopper Superchip

~456 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of ERNIE 4.5 21B A3B Thinking?

At Q4_K_M, the download is about 13.10 GB. The full-precision BF16 version is 43.65 GB. The smallest option (IQ2_XXS) is 6.00 GB.

Which GPUs can run ERNIE 4.5 21B A3B Thinking?

26 consumer GPUs can run ERNIE 4.5 21B A3B Thinking at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, NVIDIA GeForce RTX 3090 Ti, AMD Radeon RX 6800. 7 GPUs have plenty of headroom for comfortable inference.

Which devices can run ERNIE 4.5 21B A3B Thinking?

124 devices with unified memory can run ERNIE 4.5 21B A3B Thinking at Q4_K_M (13.5 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.