Baidu (ERNIE) analysis
Thesis
Baidu's public footprint reads as two tightly-coupled engines: the open-weight ERNIE 4.5 model family (dense 0.3B through a 424B-A47B MoE, with vision-language and "Thinking" reasoning variants) shipped on Hugging Face, and the PaddlePaddle open-source stack that serves as its distribution and tooling moat — led by PaddleOCR, by far its largest community asset. Document AI / OCR is the clear commercial center of gravity: its single most-downloaded model is an OCR model, and its most-active release cadence is on PaddleOCR/PaddleX.
Shipping
Models (Hugging Face, 30-day downloads):
- `baidu/Qianfan-OCR` — 204,259 downloads, 1,177 likes (~4.7B params), the most-downloaded model in the set; OCR is the demand center.
- `baidu/ERNIE-4.5-VL-28B-A3B-PT` — 202,248 downloads; the most-pulled vision-language MoE (~29B params, 3B active).
- `baidu/ERNIE-4.5-21B-A3B-PT` — 66,235 downloads; the workhorse mid-size MoE.
- `baidu/ERNIE-Image` — 35,776 downloads, 641 likes; image generation.
- `baidu/ERNIE-4.5-0.3B-PT` — 29,297 downloads; the small dense edge model.
- Reasoning push: `baidu/ERNIE-4.5-21B-A3B-Thinking` (12,856 downloads, 786 likes) and `baidu/ERNIE-4.5-VL-28B-A3B-Thinking` (539 likes) — high like-to-download ratios suggest strong interest relative to actual pull.
- Frontier scale, low adoption: `baidu/ERNIE-4.5-300B-A47B-PT` (~300B params, 1,375 downloads) and `baidu/ERNIE-4.5-VL-424B-A47B-Base-PT` (~424B params, 142 downloads) — the big MoEs are published but lightly pulled, consistent with their cost.
The lineup spans a full ladder — 0.3B dense → 21B/28B A3B MoE → 300B/424B A47B MoE — in PT, Base-PT, and Thinking flavors, signaling a deliberate one-family-many-sizes strategy. It also includes a Qianfan VL line (`baidu/Qianfan-VL-8B`) and image variants (`baidu/ERNIE-Image-Turbo`, `baidu/ERNIE-Image-Aes`).
Repos (GitHub stars):
- `PaddlePaddle/PaddleOCR` — 81,367 stars, the dominant asset (PaddleOCR shipped v3.4.1 → v3.6.0 in the recent window).
- `PaddlePaddle/Paddle` — 23,938 stars, the core framework.
- `PaddlePaddle/PaddleDetection` (14,238), `PaddlePaddle/PaddleFormers` (12,982), `PaddlePaddle/PaddleNLP` (12,952), `PaddlePaddle/PaddleSpeech` (12,614), `PaddlePaddle/ERNIE` (7,722), `PaddlePaddle/PaddleX` (6,156).
Releases: the recent cadence is heavily PaddleX/PaddleOCR — PaddleOCR v3.6.0, PaddleX v3.6.1 and v3.5.0–v3.6.0, plus FastDeploy v2.5.0 (inference deployment) and PaddleFormers 1.1.1 — i.e. shipping is concentrated on the OCR product surface and the deployment/training tooling around it.
Research themes
No first-party writing captured yet.
Hiring & scaling
Open roles are entirely US-based (Mountain View and Sunnyvale, CA) and split into two clear buckets:
- Silicon / ML systems in Sunnyvale: *Machine Learning System Hardware Architect*, *Machine Learning System Software Architect*, *CPU Digital Design Engineer*, *CPU/GPU/Processor Hardware Architect*, and *Design Verification Engineer*. This is a hardware-and-systems hiring cluster — chip-level design and ML-systems architecture, pointing at in-house compute/accelerator investment rather than model research.
- Commercial / GTM in Mountain View: *GTM Strategy & Operations (AI Desktop & Mobile App)*, *Senior Business Development & Partnerships Manager*, *Advertising Sales Manager*, multiple *Account Manager* and sales/campaign roles, and *Head of Global Business Unit, Baidu USA*. The "AI Desktop & Mobile App" GTM role plus the ad-sales and BD weighting signal a US consumer-app monetization and partnerships push.
There is also an *R&D Software Engineer Lead & CISO* (Mountain View). Notably, the captured roles contain no model-research/applied-science postings — the visible US hiring is hardware-systems + go-to-market, not core model R&D.
Traction highlights
- Most-starred repo: `PaddlePaddle/PaddleOCR` at 81,367 stars — the standout, several times larger than the core Paddle framework (23,938).
- Most-downloaded model: `baidu/Qianfan-OCR` at 204,259 30-day downloads, narrowly ahead of `ERNIE-4.5-VL-28B-A3B-PT` (202,248).
- Hacker News: thin traction — the top thread is `PaddlePaddle/PaddleOCR` at 19 points / 3 comments, with `PaddlePaddle/PaddleSeg` at just 1 point. Baidu's reach shows up in download/star counts, not in HN discussion.
- Most-liked models skew toward the reasoning variants — `ERNIE-4.5-21B-A3B-Thinking` (786 likes) and `Qianfan-OCR` (1,177 likes) lead community attention.