Baidu released a giant model you cannot talk to yet, and that is the whole point

Cover image: AI-generated

Hello there,

A giant model you cannot talk to yet.

That sounds wrong at first. Baidu released ERNIE 5.5 Flash Base on November 5, and the weights are out there to download. So why can you not just open it and chat?

Well, not quite. It is open, but it is unfinished on purpose.

This is a base model. It stopped at the end of pre-training, before the training that teaches a model to chat. Baidu trained it from scratch on 33.5 trillion tokens, and it is text only. It has 280 billion parameters in total. Guess how much of it actually wakes up for one question? Only 14 billion out of 280 billion. The rest stays quiet for that token.

Yesterday I wrote about Tencent's Hy4, a finished model that can see a screen and work inside apps. This one is the opposite kind of release. It is a starting point for other people to build on. A research team can take this checkpoint and test its own training ideas, instead of repeating Baidu's pre-training run. The weights are on Hugging Face, and Baidu says the finished chat version will follow later. No date was given.

In its own tests, the model scored 93.3 on MMLU and 82.3 on HumanEval+, ahead of DeepSeek V4 Flash Base on those two. But it was not a clean sweep. Hy3 preview led on FullStackBench, and DeepSeek led on InfiniteBench. Those are Baidu's own figures, so I read them as a claim to check.

The licence has limits too. It leaves out the EU, the UK and South Korea, and big commercial use needs Baidu to say yes first. So the word open here is narrower than it sounds.

I like one thing about this release. Baidu did not dress it up as a finished product. It said, here is the raw part, now you do the rest. That is honest. A base model is only as good as what people build from it.

Thanks for reading. I will watch for the finished version, and I will tell you what it can actually do when it lands.