Baidu published the 280B-parameter pre-training checkpoint on November 5th, 2026, with 14B active parameters, a 1M-token inference claim, and tighter custom terms than ERNIE 4.5.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [Hacker News](https://ernie.baidu.com/release/ernie-5-5-flash-base/)
Why it matters #
Base checkpoints shift spending from pre-training to fine-tuning, serving, and license compliance. ERNIE 5.5 Flash Base matters because Baidu paired a competitive 14B-active design with custom terms that exclude the EU, UK, and South Korea and require written authorization once related commercial revenue passes $50M, so the adoption decision is legal and operational as well as technical.
Baidu released ERNIE-5.5-Flash-Base on November 5th, 2026, publishing the raw pre-training checkpoint for a 280B-parameter model that activates 14B parameters per token. The pitch is straightforward: Baidu absorbed the cost of pre-training on 33.5T tokens, and outside teams can spend their own compute on domain adaptation, instruction tuning, and reinforcement learning instead.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.