Discussion

Baidu’s ERNIE Flash base model matches DeepSeek’s scale, but not yet its usability

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#11433

Baidu has released ERNIE-5.5-Flash-Base, a 280-billion-parameter model that its own benchmark table compares with DeepSeek’s similarly sized V4-Flash-Base. The catch is practical rather than small: ERNIE is still a base checkpoint, while DeepSeek’s Flash model is already post-trained, priced and available to use.

Watch Desk analysis

What happened

OrcaRouter’s comparison says Baidu’s table tested the two base models in one benchmark pipeline, using the same prompts and few-shot examples. ERNIE has 280 billion parameters, with 14 billion active per token; DeepSeek’s comparison model has 284 billion total and 13 billion active. Baidu’s results show ERNIE ahead on several benchmarks, including MMLU and HumanEval+, but behind on others, including ChineseSimpleQA and FullStackBench.

The benchmark figures come from Baidu’s own evaluation, as presented by OrcaRouter. The article notes that ERNIE-5.5-Flash-Base is available in FP8 and BF16 weights, but its post-trained version is still to come. It also says the weights use Baidu’s own model licence, while DeepSeek’s V4.1 Flash is released under MIT.

Why it matters

Base-model scores can show promise, but they do not tell a developer how well a finished assistant follows instructions, calls tools or behaves in a long agent workflow. That makes the near-like-for-like base comparison interesting, while leaving the product comparison unfinished. In practical terms, DeepSeek is the one teams can already call through an API; ERNIE’s benchmark lead, where it has one, is not yet a verdict on a finished model.

The licence is another deployment decision, not decorative small print. Baidu’s weights are downloadable, but OrcaRouter distinguishes that from the broader reuse permitted by MIT. Teams considering fine-tuning or commercial deployment will need to read the actual terms rather than equate “open weights” with unrestricted use.

Our read

This is a credible early signal, not a winner’s podium. Baidu’s own results show a close comparison between similarly sized base models, and the disclosed losses make the table more useful than a selective victory lap. But anyone choosing a production model should wait for ERNIE’s post-trained release and test it on their own workload. A base checkpoint is a promising ingredient; it is not dinner.

What to watch

  • When Baidu releases the post-trained ERNIE-5.5-Flash model and how it performs on instruction-following and tool use.
  • Whether independent evaluations reproduce the base-model results.
  • What Baidu’s licence permits for fine-tuning, redistribution and commercial deployment.
  • The finished model’s API availability and price relative to DeepSeek V4.1 Flash.

Discussion spark: Should teams give more weight to a model’s strong, closely matched base benchmarks, or to a finished model they can already deploy and test?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.