Baidu releases ERNIE 5.5 Flash Preview: 280B/14B MoE with 1M context, RMB 1/4 per million tokens, Apache 2.0 weights promised by Nov 25

Why it matters

【Key points】

· Launch: released Nov 11 on Baidu's ERNIE site and announced at Baidu World on Nov 12; the first post-trained model on ERNIE 5.5 Flash Base (Nov 5).

· Specs: text-only MoE, 280B total / 14B active parameters, 1M-token context.

· Pricing: API live on Qianfan at RMB 1 per million input tokens, RMB 4 output, RMB 0.04 cached input. Free on OpenRouter as baidu/ernie-5.5-flash-preview:free.

· Weights: to be open-sourced under Apache 2.0, by Nov 25 per the press release. Only the Base model is on Hugging Face so far.

· Stage: an intermediate post-training checkpoint. Baidu credits ERNIE Evolve, its in-house recursive self-improvement tech, for efficiency gains, without details.

· Self-reported: Terminal-Bench 4.0 46.5, AutomationBench 61.9, DeepSWE v1.1 71.2, HLE (no tools, text-only) 46.4.

【Our read】

Background

API first, open weights about two weeks later. Baidu also launched Token Exchange, a model-service marketplace.

Comparison

• 14B active is smaller than MiniMax M3.1 Flash (~428B/23B) or Ling-3.1-flash (560B/25B).
• At roughly $0.14/$0.57 it matches MiniMax ($0.15/$0.60) and costs about half as much as DeepSeek V4.1 Flash ($0.30/$1.20).
• Baidu's own table puts it ahead of DeepSeek-V4.1-Flash-1017 (42.3) on Terminal-Bench but behind it on DeepSWE (77.0).
• The Base license excludes the EU, UK and South Korea; Apache 2.0 for Preview would be a real change.

Impact

Flash-tier prices go lower; the Apache 2.0 weights, benchmarks and RSI claim remain unverified.

【Verifiable】

• · ERNIE release page (2026-11-11)
• · Baidu press release (2026-11-12)
• · OpenRouter
• · Base license