← Analysis archivereport.jsonevidence.jsonDossier →
Baidu (ERNIE)Frontier labgenerated Jun 27, 2026 · 4w

Baidu (ERNIE) analysis

15 cited sources485 signals on file

Thesis: Baidu is on a dual track — shipping competitive frontier models (ERNIE 5.1 at claimed 6% training cost of peers, Unlimited-OCR, NAVA) while aggressively building US enterprise sales/GTM teams and a Sunnyvale silicon design org. The formation of a centralized Model Committee W5 signals internal R&D consolidation as the company pivots toward the "agent era."

Key signals:

  • Hiring (~20 roles cited): Heavy GTM/sales buildout in Mountain View P5E14E21, a dedicated silicon design team in Sunnyvale for Kunlunxin chips E47E49, and Canada market entry E43E48
  • Releases: PP-OCRv6 (surpasses VLMs at 34.5M params) P6, Unlimited-OCR (1.8K GitHub stars in <24h) W2E2, PaddleFleet v0.3.0 + PaddleFormers v1.2.0 shipped simultaneously E12E11
  • Talking: Systematic LMArena benchmark campaigns across 5 ERNIE versions [E32](/blog/posts/ernie-5.1-0508-release/)[E36](/blog/posts/ernie-5.1-preview-0430-release-on-lmarena/)[E60](/blog/posts/ernie-5.0-preview-1022-release-on-lmarena/), "Daily Active Agents" as new success metric W6, ERNIE 5.1's cost-efficiency narrative W3
  • Thin spots: No fork activity cited, no safety evidence, no core ML research hiring in this pack, and ERNIE 5.1 remains closed-weight/platform-gated W3