baidu

/

ERNIE-4.5-21B-A3B-Thinking

like 759

Follow

BAIDU 1.44k

Is it possible

I strongly suspect that some people are artificially boosting the download counts and trending status of Baidu’s models on Hugging Face to create hype for their domestic marketing. This kind of manipulation actually harms the open-source community.

As Chinese models, users typically choose modelscope.cn as an alternative to huggingface.co, so their download numbers should be closely related.

But here’s what the data shows:

• ERNIE-4.5-21B-A3B-Thinking: 100k downloads on Hugging Face, only 670 downloads on ModelScope, less than 40M token usage in OpenRouter per day and just 50 code snippets on GitHub.
• Qwen3-Next-80B-A3B-Instruct: 90k downloads on Hugging Face, 18.4k downloads on ModelScope, more than 1B token usage in OpenRouter per day and 416 code snippets on GitHub.

Another sign is that ERNIE-4.5-21B-A3B-Thinking gets far less discussion on communities like Reddit compared to the Qwen3 series models.

Something doesn’t add up.

Is it possible that most downloaders of this compact model are western enterprises or users that can easily run the model themselves, and China users mainly use the model conveniently on Baidu's Qianfan platform? Maybe try it for yourself?

Yea I was gonna say...

Is it possible that most downloaders of this compact model are western enterprises or users that can easily run the model themselves, and China users mainly use the model conveniently on Baidu's Qianfan platform? Maybe try it for yourself?

Huge difference b/w the 2 models. ERNIE is a 21B model with quants for all sorts of different hardware, while Qwen3-Next has 80B params and only bnb/mlx quants.

Let @unsloth cook and you'll see the numbers start to change.

But what is looks so fake is that it's so close to Gemini 2.5 pro
Like Qwen3-Next-80B-A3B-Instruct is nearly something like Gemini flash 2.5
But this one it's 20B but it's something close to Gemini 2.5 pro just feels like over marketing
I mean they could say like flash 2.5 i would say proply but pro it's just fake

We are all technicians, don't jump to conclusions based on speculation, check more

this model can't even answer a simple "hi" message on unsloth's 8 bits on ollama.

Is it possible

I strongly suspect that some people are artificially boosting the download counts and trending status of Baidu’s models on Hugging Face to create hype for their domestic marketing. This kind of manipulation actually harms the open-source community.

As Chinese models, users typically choose modelscope.cn as an alternative to huggingface.co, so their download numbers should be closely related.

But here’s what the data shows:

• ERNIE-4.5-21B-A3B-Thinking: 100k downloads on Hugging Face, only 670 downloads on ModelScope, less than 40M token usage in OpenRouter per day and just 50 code snippets on GitHub.
• Qwen3-Next-80B-A3B-Instruct: 90k downloads on Hugging Face, 18.4k downloads on ModelScope, more than 1B token usage in OpenRouter per day and 416 code snippets on GitHub.

Another sign is that ERNIE-4.5-21B-A3B-Thinking gets far less discussion on communities like Reddit compared to the Qwen3 series models.

Something doesn’t add up.

Is it possible that most downloaders of this compact model are western enterprises or users that can easily run the model themselves, and China users mainly use the model conveniently on Baidu's Qianfan platform? Maybe try it for yourself?

Unfortunately, Baidu's Qianfan platform does not support ERNIE-4.5-21B-A3B-Thinking yet.

https://cloud.baidu.com/doc/WENXINWORKSHOP/s/flxu4ej5u

I think it's unlikely this is fake - by open weighting this model it makes it pretty easy to attempt to replicate the scores or run local testing and you would find out pretty quickly that it's fake benchmarks if it were. To respond to some comments here: 1. It's not 'close to Gemini 2.5 pro' in the provided benchmarks ERNIE 4.5 scores 78.02 on AIME2025 and Gemini scores 90.05. That's a pretty big difference, and we've seen opensource small models beat giant closed source ones multiple times before. 2. And to reply to the original message that started this discussion: The download count is NOT necessarily related to the download count on other platforms for two reasons -1 The download difference might simply be because of the popularity and accessibility of huggingface, I know systems like lmstudio default to pulling models from this site. 2- the types of people who visit these different platforma may be looking for various model types - and, for example - might be looking for more general purpose models on huggingface, and possibly other related platforms are mainly for some other model form like mathematics based.

Even through all of that I'm not saying that this model is 100% not faked, I'm just saying some of your points could be explained

All in all I would indeed trust the results of Qwen3-next or GPT-oss 20B more than this model, but still I wouldn't expect the benchmarks or download count to be very tampered with

By the way, Can instruct model(non-think) like qwen3-80b-a3b-next（and Qwen3-xxx-2507 models） achieve the high aime2025 score？

why would i suddenly use another platform to download my models just because its closer to the country of origin?
I don't know about the download figures, but that people suddenly use modelscope because the model is chinese makes NO sense.

Additional thoughts: Huggingface would face trouble if only non chinese (open source) models would be hosted. For the better or worse that is a fact.
And 2nd why on earth would i use a chinese website if i can possibly avoid it? (presuming as always more or less state controlled)

and beside all that the busywork of managing another flipping account, settings, favorite links and stuff.... No sense!