Baidu releases ERNIE 5.5 Flash Preview: 280B params, 14B active, beats DeepSeek on 14 benchmarks
智东西 · wechat · 2026-11-11
Baidu released ERNIE 5.5 Flash Preview, the first post-trained version of its ERNIE-5.5-Flash-Base: 280B total params with 14B active, 1M context, 200+ tokens/s generation, with weights to be open-sourced by Nov 25.
- Beats DeepSeek-V4.1-Flash-1017 on 14 of 20 benchmarks; first on MCP-Atlas tool-use eval, second to Hy4 on BrowseComp and DRACO agent evals.
- First model to use Baidu's RSI technique with the ERNIE Evolve auto-research system optimizing inference. As a text-only model it iterates by converting visual issues into measurable code tests rather than viewing screens.
- Demos: built a UE5.8 xianxia game in C++ over 11 hours; two agents built airport and aircraft models in Blender and rendered videos; a FreeCAD engine CAD model; a 116-second 2D animation spanning 20 art styles.
- Time-tiered pricing at 50% off through Nov 30: peak input 1 yuan/M tokens, output 4 yuan; off-peak 0.5/2 yuan—about half of DeepSeek's price.
Baidu calls it an intermediate post-training checkpoint; the industry is moving to tiered flagship+Flash model dispatch.
Related event: Baidu Releases ERNIE 5.5 Flash Preview, Open-Sourcing November 25(9 posts)→
More from Models
- Codex safety filter flags developer's hobby git forge app as cybersecurity risk — brianmichel · 2026-11-11
- DeepSeek's open-source coding agent DSH statistically ties for first on HF Harness Leaderboard — TheZachMueller · 2026-11-11
- US open-source models will dominate thanks to profitable business models — marktenenholtz · 2026-11-11
- NVIDIA open-sources GR00T 2 robot model running real dexterous chem lab demos — EricBuess · 2026-11-11
- Price cut, bill up 31%: new model's self-checking eats Anthropic's token discount — tengyanAI · 2026-11-11
- US data vendors accused of selling RL environments to Chinese military-linked firms — xeophon · 2026-11-11