Baidu launches ERNIE 5.5 Flash: 280B params, 14B active, coding-first, soon open source

量子位 · wechat · 2026-11-11

Baidu unveiled ERNIE 5.5 Flash Preview a day before its World Conference, promising an Apache 2.0 open-source release with SGLang/vLLM support. Specs: 280B total parameters with 14B active per inference, 1M-token context, and 200+ tokens/sec generation, with coding and agent capabilities as the headline focus.

Showcase demos: a 116-second pelican-on-a-bicycle animation spanning 20 art styles via Python code, and an 11-hour UE5.8 run that built a playable wuxia game
Hands-on tests: a single-HTML playable Backrooms game with tape-collection and chase mechanics, and a Suzhou garden modeled from scratch in Blender in 30 minutes
End-to-end training recipe: 14M executable code tasks that must actually run, multi-agent trajectories teaching tool use and error recovery, and converting visual bugs into code-checkable data for self-debugging
Per-scenario skills: UE/Unity/Godot, Blender/CAD, 1000+ creation skills for video; hybrid sparse attention, multi-router MoE, and FP4 expert weights for speed
First use of in-house RSI, with ERNIE Evolve optimizing the model's own inference; API priced at a limited-time 50% discount on Qianfan

Original post →

Baidu launches ERNIE 5.5 Flash: 280B params, 14B active, coding-first, soon open source

More from coding & agent