Baidu releases ERNIE 5.5 Flash Base
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Baidu released ERNIE 5.5 Flash Base, a 280B-parameter sparse MoE model with 14B activated per token.
It is a pretraining starting point, not a chatbot, under a restrictive licence.
Background
Baidu is the Chinese company whose ERNIE Team announced the release. ERNIE 5.5 Flash Base is a sparse MoE (mixture-of-experts) base language model: 280B total parameters, of which 14B are activated per token, pretrained from scratch on 33.5T tokens. Baidu says the weights are the end-of-pretraining checkpoint, before any mid- or post-training, meant as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research, and to spare developers the cost of large-scale pre-training. It marks a third licensing position for Baidu after ERNIE 4.5's permissive Apache 2.0 and ERNIE 5.0's release with no weights at all.
How it unfolded
- Jun 30, 2025Baidu released ERNIE 4.5 with ten variants under Apache 2.0, described as genuinely permissive with no territory, revenue gate or attribution string.
- Nov 2025Baidu released ERNIE 5.0 and shipped no weights at all.
- Oct 21, 2026The baidu/ERNIE-5.5-Flash-Base repository was created on Hugging Face.
- Nov 2, 2026 -3, 2026The repository took its 128 FP8 weight shards.
- Nov 4, 2026Architecture and benchmark diagrams landed.
- Nov 5, 2026Baidu officially open-sourced ERNIE 5.5 Flash Base, pretrained on 33.5T tokens with a training context of up to 128K and support for a 1M-token context in fine-tuning and inference; weights went to Hugging Face in FP8 and BF16.
- Nov 5, 2026Baidu's release page was dated that day, but the model card was still being revised that morning, with no press release, launch event or press coverage to speak of.
Key terms
- sparse MoE
- A machine learning technique using multiple expert networks to divide a problem space into regions.
- Apache 2.0
- A permissive free-software license from the Apache Software Foundation allowing use, modification, and redistribution without royalties.
Source: Wikipedia
Who’s saying what
- Official
- Baidu says a base model's value lies both in the capability foundation pretraining establishes and in the starting point it gives for later training, and that a post-trained ERNIE 5.5 Flash model will follow.
- Analysts
- OrcaRouter calls the release quiet, notes ERNIE-5.5-Flash-Base and Qwen 3.8 Flash are not the same kind of object — one a starting point you still have to train, the other a product you can call — and says the benchmark table cannot be dropped into a leaderboard next to chat models.
- Caution
- OrcaRouter says the release's practical reach is bounded by the hardware needed for full-model fine-tuning and a licence that restricts geography, larger-scale commercial use and transfers of capability to unrelated models.
Still unverified
The Hugging Face sidebar rounds total parameters to 286B, while the model card says both figures exclude 0.36B of what Baidu calls ECHO parameters. Baidu's benchmark comparison sets its LiveCodeBench number against Qwen's 40.1 and 91.9, which one source calls a trap because the two are not comparable.