Baidu releases ERNIE 5.5 Flash Preview: 280B MoE, 14B active, 200 tok/s
赛博禅心 · wechat · 2026-11-11
Baidu released ERNIE 5.5 Flash Preview, the first post-trained version of last week's open-sourced Base model.
Specs: text-only MoE, 280B total / 14B active params, 1M context, 200+ tokens/s output. API live on Qianfan (ernie-5.5-flash-preview): ¥2/¥1 per million input tokens (peak/off-peak), ¥8/¥4 output, limited-time 50% off. Weights open-source on Nov 25.
Demos:
- Terminal-Bench 4.0: 46.5, ahead of several Chinese rivals but behind frontier models
- Built a 3D wuxia game in C++ in 11 hours by converting playtest issues into measurable code tests
- Two agents co-built airport + plane in Blender; recreated Horyu-ji temple; made 2,100-object whitebox previz for film; editable FreeCAD W16 engine and lunar rover
- Delivered research reports, Excel with Monte Carlo simulation, an original symphony
- Wrote a 116-second vector-animated peloton-on-bicycle film, 10k+ lines of code
Architecture: 512 experts/layer picking 10 with multi-router scoring; 37 of 43 layers compress every 128 tokens into one cache entry; ECHO speculative decoding. Part of inference optimization was done automatically by ERNIE Evolve, Baidu's RSI system, which beat public bests on NanoChat Auto Research and NanoGPT Speedrun.
Related event: Baidu Releases ERNIE 5.5 Flash Preview, Open-Sourcing November 25(9 posts)→
More from Infra
- Price cut, bill up 31%: new model's self-checking eats Anthropic's token discount — tengyanAI · 2026-11-11
- Anthropic IPO demand so strong bankers weigh higher price at $1.4T valuation — firstadopter · 2026-11-11
- SimpleHarn Routes Local LLMs and Refuses Loads That Exceed Your VRAM — legislat · 2026-11-11
- Rivet launches SQLite Forking: 20ms agent-thread forks with zero data copying — samgoodwin89 · 2026-11-11
- French netizen proposes datacenters pay local residents compensation — clemnt · 2026-11-11
- GPU Rental Prices Hit Record Highs: B200 at $6.97/hr, H100 at $3.06 — inductionheads · 2026-11-11