← November 6, 2026 briefing · added in the 12:00 KST update

P2 Open source Open weights Language modelsFollow-up12:00 updateBaidu

ERNIE 5.5 Flash Base open-source follow-up: 280B-total, 14B-active MoE, 128K training context, up to 1M tokens supported (report)

Verified (fact)

According to Pandaily, ERNIE-5.5-Flash-Base, released by Baidu, is a raw checkpoint after pretraining. It is an MoE model with 280B total parameters and 14B active parameters per token. The training context length was 128K, and it supports up to 1 million tokens. It was released for use in further training.

Why it matters (interpretation)

The scale of a model whose release had previously been announced without details is now confirmed. Teams building their own models gain another option: a large MoE base model with 14B active parameters that can be fine-tuned or trained with reinforcement learning directly.

Confidence medium · official source pending

Sources

Story thread

  1. 2026-11-05 Baidu open-sources base model ERNIE 5.5 Flash Base, claims best-in-class results on several benchmarks (report)
  2. 2026-11-06 ERNIE 5.5 Flash Base open-source follow-up: 280B-total, 14B-active MoE, 128K training context, up to 1M tokens supported (report)
  3. 2026-11-07 ERNIE 5.5 Flash Base follow-up: weights license excludes South Korea and Europe from permitted regions (report)

More about Baidu