ERNIE 4.5 VL 28B A3B PT — Hardware Requirements & GPU Compatibility
VisionERNIE 4.5 VL 28B A3B is Baidu's multimodal mixture-of-experts chat model, featuring 28 billion total parameters and only 3 billion active per token. It processes both text and images, supports thinking as well as non-thinking modes, and relies on a heterogeneous MoE design with distinct text and vision experts (64 each, 6 activated, plus 2 shared). The PT suffix indicates PyTorch Transformers weights rather than the PaddlePaddle build. Since just 3 billion parameters activate during inference, decoding stays fast; however, all weights still need to be stored, meaning it fits on a single consumer GPU with roughly 24 GB of VRAM once quantized. The model handles a context window of up to 131,072 tokens. Released under the Apache 2.0 license, it allows unrestricted commercial and research use. Published in June 2025, this model joins the ERNIE 4.5 family, whose documentation also lists larger A47B MoE variants.
ERNIE 4.5 VL 28B A3B PT runs on a 64 GB+ Mac or a workstation GPU
Best local pick → Mac Pro M2 Ultra (64 GB) · BF16 · 60 GB · ~61 tok/s
Macs from 64 GB (BF16 · ~61 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run ERNIE 4.5 VL 28B A3B PT
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.
Recommended
vllm serve baidu/ERNIE-4.5-VL-28B-A3B-PT --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on A100 80GB (20 GB headroom) · BF16
- Generation speed
- ~164 tok/s
- generation speed
- Cost per 1M output tokens
- $1.93
- per 1M output tokens
How Much VRAM Does ERNIE 4.5 VL 28B A3B PT Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 59.2 GB | 61 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 60 GB |
Selected BF16
Brain floating point 1660.0GBat 16K context
- Weights
- 58.8 GB
- KV cache
- 0.9 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 36.0 GB more
Or .
Which Hardware Can Run ERNIE 4.5 VL 28B A3B PT?
BF16 · 59.2 GBERNIE 4.5 VL 28B A3B PT (BF16) requires 59.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 77+ GB is recommended. Using the full 131K context window can add up to 7.4 GB, bringing total usage to 66.6 GB. No consumer GPU has enough memory. 56 devices with unified memory can run ERNIE 4.5 VL 28B A3B PT, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), Mac Pro M2 Ultra (64 GB).
64–80 GB
fits · 4.0 GB free · up to 64K ctx
3 Mac classes · 1 device family
M2 Ultra*64 GB~61Simulate output on Mac Pro M2 Ultra (64 GB)M5 Max64 GB~51Simulate output on Mac Studio M5 Max (64 GB)Orin~21Simulate output on NVIDIA Jetson AGX Orin 64GB
By VRAM tier · typical GPUs and Macs at BF16
Alternatives to ERNIE 4.5 VL 28B A3B PT
ERNIE family · 7 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.8 27BAlibaba | 121.9 | 17.9 GB | ~38 | same size |
| Hy Memory 30B A3BTencent | — | 19.1 GB | ~184 | same size |
| Granite Switch 4.1 30BIBM | — | 22.7 GB | ~35 | same size |
| Muse Glimmer 30BMeta | — | 19.6 GB | ~36 | same size |
| NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA | — | 19.3 GB | ~197 | same size |
| Hy MT2 30B A3BTencent | — | 19.1 GB | ~170 | same size |
Specifications
- Publisher
- Baidu
- Family
- ERNIE
- Parameters
- 29.4B
- Active parameters
- 3B per token
- Architecture
- Ernie4_5_VLMoeForConditionalGeneration
- Context Length
- 131,072 tokens
- Vocabulary Size
- 103,424
- Release Date
- 2025-06-28
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does ERNIE 4.5 VL 28B A3B PT need?
ERNIE 4.5 VL 28B A3B PT requires 59.2 GB of VRAM at BF16. Full 131K context adds up to 7.4 GB (66.6 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 29.4B × 16 bits ÷ 8 = 58.8 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
Can NVIDIA GeForce RTX 5090 run ERNIE 4.5 VL 28B A3B PT?
No — ERNIE 4.5 VL 28B A3B PT requires at least 59.2 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run ERNIE 4.5 VL 28B A3B PT on a Mac?
Yes — Mac Pro M2 Ultra (64 GB) and 44 other Macs can run ERNIE 4.5 VL 28B A3B PT. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 59.2 GB of usable unified memory (RAM minus macOS overhead).
Can I run ERNIE 4.5 VL 28B A3B PT locally?
Yes — ERNIE 4.5 VL 28B A3B PT can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is ERNIE 4.5 VL 28B A3B PT?
At BF16, ERNIE 4.5 VL 28B A3B PT can reach ~146 tok/s on AMD Instinct MI350X. 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 → 6.0 GB active ÷ (8000 × 0.65) = 1.16 ms, plus 28 layers × 0.055 ms = 1.54 ms, so 1000 ÷ 2.70 ms = ~370 tok/s
Estimated speed at BF16 (59.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of ERNIE 4.5 VL 28B A3B PT?
At BF16, the download is about 58.80 GB.
Which GPUs can run ERNIE 4.5 VL 28B A3B PT?
No single consumer GPU has enough VRAM to run ERNIE 4.5 VL 28B A3B PT at BF16 (59.2 GB). Multi-GPU or professional hardware is required.
Which devices can run ERNIE 4.5 VL 28B A3B PT?
56 devices with unified memory can run ERNIE 4.5 VL 28B A3B PT at BF16 (59.2 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (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.