FULL STORY

Baidu's ERNIE 5.5 Flash: From Open-Source Base to Preview

Baidu open-sourced the ERNIE-5.5-Flash-Base MoE model (280B total, ~14B active, 1M context) on Nov 5, then followed up on Nov 11 with its first post-trained preview version.

2026-11-05 ~ 2026-11-11 · 2 episodes · 31 posts

Episode 1 · Baidu Open-Sources ERNIE 5.5 Flash Base: 280B MoE with 14B Active Params and 1M Context (2026-11-05, 22 posts)

Baidu officially released and open-sourced ERNIE-5.5-Flash-Base on Nov 5: an efficient sparse-MoE base model with 280B total parameters, only 14B (5%) activated per token, pre-trained from scratch on 33.5T tokens, supporting up to 1M-token context for fine-tuning and inference. Weights are available on Hugging Face with Day-0 vLLM support; official benchmarks claim SOTA among same-size models, and it hit the HF trending chart on Nov 6. This is a rare case of a major lab shipping a true "base" model and another heavyweight open-source practice of the sparse MoE route.

Confirmed

  • Sparse MoE architecture: 280B total / 14B active per token; per @teortaxesTex, the backbone has 43 layers—1 Dense FFN layer and 42 Multi-Router Latent MoE layers—with hybrid sparse attention
  • Pre-trained from scratch on 33.5T tokens; per the official blog, this is a raw base checkpoint with no mid- or post-training, leaving post-training directions to developers; a full ERNIE 5.5 Flash with post-training will follow; @机器之心 reported 128K training context extrapolating higher for fine-tuning/inference, positioned for high-frequency long-horizon Agent workloads
  • Up to 1M-token context for fine-tuning and inference; @karminski3, after reading the model card and config, argues the model was designed from scratch for "1M ultra-long context + post-training/Agent repeated-sampling efficiency"
  • Weights on Hugging Face, including a BF16 version (286B each), in safetensors format with two arXiv tech reports; the Collection also includes Unlimited-OCR, ERNIE-Image, ERNIE 4.5, and Qianfan; release notes at ernie.baidu.com plus an official blog detailing the attention and MoE design
  • Official benchmarks claim SOTA across knowledge, reasoning, math, code, multilingual and long-context tasks among same-size models; @karminski3 notes it ranks first across 25 base-model tests, compared against DeepSeek-V4-Flash-Base and Hy3-preview-Base
  • The release coincided with the World Congress 2026 announcement of open weights for the ERNIE 5.5 family; vLLM announced Day-0 support, and Baidu is working with SGLang and other communities on ecosystem support
  • Per @BaiduInc, this is the first release by returning Baidu employee tianxiangsun (Sun Tianxiang)

Why it matters

  • The 280B-total / 14B-active design balances capacity with low inference cost, another major open-source milestone for sparse MoE
  • Million-token context and open weights give the community a fresh starting point for fine-tuning, long-document, and long-horizon Agent applications; Day-0 vLLM support and the SGLang ecosystem lower adoption barriers
  • @XianbaoQIAN notes this is a rare big-lab release of a true "base" model—developers can fine-tune, align, or apply RL on a clean base without inheriting official post-training trade-offs; attention from outside observers like @teortaxesTex on the Multi-Router Latent MoE details shows the architecture has sparked real discussion in the technical community

2 more related posts →

Episode 2 · Baidu Releases ERNIE 5.5 Flash Preview, Open-Sourcing November 25 (2026-11-11, 9 posts)

On November 11, Baidu officially released ERNIE 5.5 Flash Preview, the latest open-weight foundation model in the ERNIE family and the first post-trained version of ERNIE-5.5-Flash-Base, which was open-sourced last week. It is also the first official model built on a new architecture designed for efficient inference.

Confirmed

  • Text-only MoE model with 280B total parameters and 14B activated per token
  • 1 million token context window, output speed exceeding 200 token/s
  • Officially claimed to be stronger in coding, 3D modeling, knowledge work, and content creation, with impressive results on benchmarks like Terminal-Bench
  • According to Zhidongxi and 赛博禅心, it outperforms DeepSeek on 14 tests
  • Available for free trial on OpenRouter
  • Scheduled to be open-sourced on November 25

Why it matters

  • This is the first official model under Baidu's new architecture, marking the ERNIE series' push toward open weights and efficient inference
  • The combination of a large context window (1M) and high output speed, along with the upcoming open-source release, could lower deployment and usage barriers for developers and directly shake up the competitive landscape of open-source models (especially as a DeepSeek rival).