← Analysis archivereport.jsonevidence.jsonDossier →
Baidu (ERNIE)Frontier labgenerated Oct 18, 2026 · 6d

Baidu (ERNIE) analysis

65 cited sources500 signals on file

Thesis

Baidu's ERNIE signal in this pack splits into two tracks. On one side, a frontier-model line is openly shipping: ERNIE began as knowledge-enhanced pretraining and grew into Baidu's flagship generative line W1; the ERNIE 4.5 family went open-weight on June 30, 2025 under Apache-2.0 across ten variants (0.3B–424B) W2, part of a 2025 open-source pivot that produced 100+ releases W3; this was followed by ERNIE 5.0 (a 2.4T-parameter multimodal MoE) W4[E52](/blog/posts/ernie5.0/) and ERNIE 5.1 (claimed at 6% of comparable pre-training cost) [E49](/blog/posts/ernie-5.1-0508-release/). On the other side, Baidu's US hiring signal is dominated by programmatic-advertising operations and sales for MediaGo's DSP rather than model research P3P4P5P6P7P8E23E37. The observable data-business surface is therefore infrastructure/serving and evals via the PaddlePaddle release train, plus ad-tech and AI-search GTM — while the lab's own ML engineering hiring is thin in this pack.

Signal desks

Hiring: Repeated account-manager, campaign-manager, and ad-ops roles for MediaGo's programmatic DSP across Mountain View, Remote US, and Toronto, with Mandarin proficiency a recurring requirement P3P4P5P6P7P8. Sales/GTM buildout includes Head of Sales – North America, Account Executive (New Business), AI Solutions B2B Sales Executive, Advertising Sales Associate, and GTM Strategy & Operations (AI Desktop & Mobile App) E20E23E37E44E45. The only AI-infrastructure-facing hiring is Forward Deployed Engineer (Intern) for the Querit Search API (AI agents/RAG) in Mountain View and Los Angeles P9E21E26. No cited model/evals/data/safety research roles appear in this pack.

Forks: No cited fork events (no forked upstream parent repos). The nearest evidence is first-party PaddlePaddle repos appearing: FlashMask (mask-aware distributed attention for long-context training) E8P2, VHA (NeurIPS 2026 Virtual Head Attention for efficient LLM architecture and training) E9P1, PaddleFleet (distributed training library) E31, GraphNet (computation-graph database for tensor compiler research) E30, plus hardware-port repos (PassNet, paddle_musa, Paddle-Metax, TST, Paddle-iluvatar) E32E33E34E35E36.

Releases: Hugging Face model releases span OCR, VL, thinking, image, and video: Unlimited-OCR, Qianfan-OCR, ERNIE-4.5-21B-A3B-Thinking, ERNIE-4.5-VL-28B-A3B-Thinking, ERNIE-Image/-Turbo/-Aes, NAVA, and Qianfan-VL-3B/8B/70B E1E2E3E4E5E6E7E10E14E19E24. The framework train is dense: Paddle v3.4.0 P10 plus iterative PaddleX/PaddleOCR/FastDeploy releases adding OCR-VL models, benchmarking, serving, and domestic-hardware support P14P16P19P23P27.

Talking: Leaderboard-led narrative — repeated LMArena posts for ERNIE 5.0 previews and ERNIE 5.1 claiming No.1 China / top global ranks [E50](/blog/posts/ernie-5.1-preview-0430-release-on-lmarena/)[E54](/blog/posts/ernie-5.0-0110-release-on-lmarena/)[E55](/blog/posts/ernie-5.0-preview-1220-release-on-lmarena/)[E56](/blog/posts/ernie-5.0-preview-1203-release-on-lmarena/)[E57](/blog/posts/ernie-5.0-preview-1103-release-on-lmarena/)[E58](/blog/posts/ernie-5.0-preview-1120-release-on-lmarena/)[E60](/blog/posts/ernie-5.0-preview-1022-release-on-lmarena/). Research/positioning posts cover ERNIE 5.0's unified 2.4T multimodal model [E52](/blog/posts/ernie5.0/), ERNIE 5.1's cost/RL story [E49](/blog/posts/ernie-5.1-0508-release/), ERNIE-Image's 8B DiT [E51](/blog/posts/ernie-image/), PaddleOCR-VL-1.5 (94.5% OmniDocBench) [E53](/blog/posts/paddleocr-vl-1.5/), and ERNIE-4.5-VL-28B-A3B-Thinking [E59](/blog/posts/ernie-4.5-vl-28b-a3b-thinking/). Earnings/web coverage stresses self-owned chips/models/cloud/apps and long-horizon investment W5W6.

Shipping

Open-weight HF releases with traction (downloads / likes): Unlimited-OCR 1,476,672 / 4,767 (MIT) E1; Qianfan-OCR 44,641 / 1,210 (Apache-2.0) E2; ERNIE-4.5-21B-A3B-Thinking 87,036 / 793 E3; ERNIE-4.5-VL-28B-A3B-Thinking 2,067 / 543 E5; ERNIE-Image 1,645 / 680 and ERNIE-Image-Turbo 3,201 / 409 E4E6; NAVA (text-to-video) 135 / 126 E7; Qianfan-VL-3B/8B/70B and ERNIE-Image-Aes at lower volume E10E14E19E24.

Infra/tooling shipped via the PaddlePaddle org: Paddle v3.4.0 (no notes published) P10; PaddleX/PaddleOCR added PP-OCRv5 multilingual models (EN +11%, Thai 82.68%, Greek 89.28%), CUDA 12 + ONNX Runtime, open-sourced high-stability serving, fine-grained benchmarking (per-layer/module latency + memory), and PaddleOCR-VL (0.9B VLM, 109 languages) P16P19P23; later releases added PaddleOCR-VL CPU/DCU/XPU inference, vLLM/SGLang plugin auto-install, and DOCX/LaTeX export P22P27P28. FastDeploy v2.1.0/v2.1.1 shipped KVCache scheduling (no kv_cache_ratio dependence), concurrency limits, min_p/top_k_top_p sampling, logprobs/return_token_ids, CUDA Graph (17–91% gains on ERNIE 4.5/Qwen3), MTP speculative decoding, WINT2 (ERNIE-4.5-300B-A47B +25.5%), and domestic-hardware deployment (Kunlunxin P800, Hygon K100-AI, Enflame S60, Iluvatar) P15P17. A dense minor-release cadence fixed OCR-VL issues, offline downloads, and CPU/GPU compatibility P11P12P13P18P20P21P24P25P26.

Research themes

  • Efficient long-context training: FlashMask and VHA (NeurIPS 2026) target mask-aware distributed attention and efficient LLM training optimization E8E9.
  • Distributed training & compiler tooling: PaddleFleet and GraphNet (tensor-compiler research graph DB) E31E30.
  • Cost-efficient frontier post-training: ERNIE 5.1 at 6% of comparable pre-training cost via disaggregated fully-async RL and scaled agentic post-training [E49](/blog/posts/ernie-5.1-0508-release/); ERNIE 5.0 is a 2.4T-param MoE activating fewer than 3% of params per token W4[E52](/blog/posts/ernie5.0/).
  • Native multimodality: ERNIE 5.0 unifies text/image/video/audio in a single autoregressive framework [E52](/blog/posts/ernie5.0/); ERNIE-Image uses a single-stream 8B DiT [E51](/blog/posts/ernie-image/); NAVA targets text-to-video E7.
  • Document-parsing VLMs: PaddleOCR-VL-0.9B pairs a NaViT-style dynamic-resolution encoder with ERNIE-4.5-0.3B for 109-language parsing P23; PaddleOCR-VL-1.5 hits 94.5% on OmniDocBench v1.5 [E53](/blog/posts/paddleocr-vl-1.5/).
  • Inference/serving optimization & heterogeneous hardware: KVCache, CUDA Graph, MTP, W4A8/INT4/WINT2 quantization, and domestic-GPU support P15.

Hiring & scaling

Baidu USA hiring is overwhelmingly commercialization: account/campaign/ad-ops and partner-ops roles for MediaGo's programmatic ad stack P3P4P5P6P7P8, plus sales/GTM leadership and account roles E20E23E27E28E37E44E45. AI-application enablement appears through Forward Deployed Engineer (Intern) roles for the Querit Search API targeting AI agents/RAG and production integration P9E21E26. Geography clusters in Mountain View, with Los Angeles, Toronto, and Remote US as secondary hubs P3P8E12E15E20E26. Mandarin proficiency is a recurring requirement across the ads roles P3P4P5P6. The pack shows no cited research scientist, ML engineering, data, eval, or safety postings — a gap in this evidence rather than proof of absence.

Data-business implications

  • Serving/infra tooling: FastDeploy's KVCache scheduler, concurrency controls, sampling options, logprobs/token returns, CUDA Graph, MTP, and W4A8/INT4/WINT2 quantization create deployment-configuration and latency/throughput telemetry needs P15P17.
  • Evals & benchmarking: PaddleX/PaddleOCR added fine-grained benchmarking (end-to-end plus per-layer/module latency and memory) — a direct hook for eval harnesses and performance-benchmark data products P16P19.
  • Heterogeneous hardware: FastDeploy/PaddleOCR-VL support for Kunlunxin, Hygon, Enflame, Iluvatar, DCU, and XPU implies multi-vendor deployment tooling and compatibility testing P15P27P28.
  • OCR/document pipelines: Unlimited-OCR (1.48M downloads), Qianfan-OCR, and PaddleOCR-VL signal demand for document-parsing data, OCR evals, and layout/structure datasets E1E2P23.
  • Search/retrieval GTM: Querit Search API forward-deployed roles imply retrieval-quality, latency, and RAG-integration tooling for agent builders P9E21E26; GTM Strategy & Operations for AI Desktop & Mobile App and AI Solutions B2B Sales signal consumer/enterprise GTM buildout E45E37.
  • Ad-tech data: MediaGo account/campaign/partner-ops roles imply ad performance measurement, targeting, compliance/appeals, and supply-side data workflows P3P4P5P8.

Traction highlights

  • Unlimited-OCR: 1,476,672 downloads and 4,767 likes E1; Qianfan-OCR: 44,641 downloads / 1,210 likes E2.
  • ERNIE-4.5-21B-A3B-Thinking: 87,036 downloads / 793 likes E3; ERNIE-Image: 680 likes E4.
  • Leaderboard claims: ERNIE-5.0-0110 No.1 China / No.8 global (1,460) [E54](/blog/posts/ernie-5.0-0110-release-on-lmarena/); ERNIE-5.1-Preview No.1 China / No.13 global [E50](/blog/posts/ernie-5.1-preview-0430-release-on-lmarena/); ERNIE-5.0-Preview-1220 sole Chinese model in LMArena Vision top 10 [E55](/blog/posts/ernie-5.0-preview-1220-release-on-lmarena/); ERNIE-5.0-Preview-1022 No.2 global Text [E60](/blog/posts/ernie-5.0-preview-1022-release-on-lmarena/).
  • Open-source momentum: 100+ open releases in 2025 following the ERNIE 4.5 open-weight decision W2W3.