ERNIE 4.5 VL 424B A47B: Baidu's Open Vision MoE in 2026

ERNIE 4.5 VL 424B A47B is more than a year old, which in this field usually means retired. It is still listed by API providers, though, and it still has a property that most of the open vision-language field lacks: it is very large, permissively licensed, and built around an architecture that treats images and text as separate citizens rather than bolting a vision encoder onto a language model.

The name spells out the shape. VL for vision-language, 424B total parameters, A47B for 47B active per token. This guide covers what that architecture does, what it takes to run, and how to decide whether a mid-2025 model is the right choice for image work in late 2026.

Where it came from

Baidu open-sourced the ERNIE 4.5 family on 30 June 2025: ten models released on Hugging Face, GitHub and Baidu's PaddlePaddle ecosystem, from a 0.3B dense model up to this one. The text-only flagship is ERNIE-4.5-300B-A47B; this is its multimodal counterpart, and there is a much smaller vision sibling at 28B total and 3B active.

The weights are under Apache 2.0, which permits commercial use. That matters more than it might seem, because Baidu's next generation went the other way: ERNIE 5.0 is served through Baidu's own platform, and no ERNIE 5 weights appear on Baidu's Hugging Face account at the time of writing. For an openly licensed ERNIE you can host yourself, the 4.5 family is still the line to look at.

What heterogeneous MoE means here

The model card describes a multimodal heterogeneous mixture-of-experts. Concretely, there are two expert pools:

• Text experts — 64 in total, 8 active per token.
• Vision experts — another 64, also 8 active per token.

Baidu calls the routing between them modality-isolated: image tokens and text tokens are sent to their own experts, while other parameters are shared so the two modalities still inform each other. Training used a router orthogonal loss and a multimodal token-balanced loss, both aimed at stopping one modality from crowding out the other during joint pre-training.

The design goal is that adding vision should not cost text quality, which is the common complaint about adapter-style vision models. Whether it achieves that on your data is something to test, but the reasoning is sound. The rest of the specification from the card: 54 layers, 64 query heads with 8 key-value heads, and a 131,072-token context. Mixture of experts explained covers the routing basics if the expert counts are unfamiliar.

The VL models support both thinking and non-thinking modes, so you can trade latency for step-by-step reasoning on harder visual questions such as charts and diagrams.

Running it yourself takes sixteen GPUs

The 47B active figure sets compute per token, but the 424B total sets memory, and that is the wall. The model card's vLLM example, for vLLM 0.11.2 or later, runs with a tensor parallel size of 16 and lists sixteen 80GB GPUs:

vllm serve baidu/ERNIE-4.5-VL-424B-A47B-PT --tensor-parallel-size 16 --trust-remote-code

The -PT suffix marks PyTorch-format weights for Transformers-style stacks; the -Paddle variants are for Baidu's own PaddlePaddle tooling. The card also claims lossless 4-bit and 2-bit quantisation for the ERNIE 4.5 MoE models, using what Baidu calls convolutional code quantisation, which would cut the hardware floor substantially, but measure quality on your own images before relying on a 2-bit build.

For most teams this is not a self-hosting candidate unless you already run multi-node inference. Best model for self-hosting covers what fits on more ordinary hardware, and the 28B sibling is the obvious ERNIE to try first if you need it local.

Through an API, the listing is narrower than the card

OpenRouter lists the model at $0.42 per million input tokens and $1.25 per million output at the time of writing. The listing gives a context of 123,000 tokens and a maximum output of 16,000, both below what the model card describes, and a knowledge cutoff of 31 March 2025. Text and images in, text out.

The output cap is the limit to watch. Sixteen thousand tokens is plenty for describing an image or extracting a table, but thinking mode spends output tokens on reasoning before the answer, so a long chain on a dense diagram can run into it. If answers truncate, switch to non-thinking mode or split the task.

Is a mid-2025 vision model still worth it

Age hurts a vision model in two specific ways. Its knowledge stops in early 2025, so it will not recognise recent interfaces, products or document formats. And the open multimodal field has moved quickly since, with newer models offering longer context and stronger reasoning at smaller sizes.

Where it can still win is bulk visual understanding on stable material, where price per image and a permissive licence matter more than the newest reasoning. Scanned forms, product photos, charts in reports and screenshots of mature software all fit. The comparison to run is in vision-language models compared.

A quick way to decide:

1. Take fifty real images from your workload with known correct answers, including the hard ones: small text, rotated scans, dense charts.
2. Run them through ERNIE 4.5 VL in both modes and through one newer vision model at a similar price.
3. Score exact-match accuracy on the extractable facts, not overall impressions, and record cost per correct answer.

If ERNIE matches the newer model on your images, its price and licence make it a reasonable default. If it trails, the result is the answer, and the model's age is the likely reason.