Latest22 Jun 2026Embedding/other
Unlimited-OCR
A 3B open OCR and document-parsing model that Baidu describes as pushing DeepSeek-OCR 'one step further'. It turns single images, multi-page documents and PDFs into structured text with layout markers and bounding boxes. It is released under the MIT licence.
- 3B parameters, MIT licence
- Single-image, multi-page and PDF parsing
- Structured output with detection boxes
- Long-horizon parsing of long documents
AccessOpen weights on Hugging Face (MIT); vLLM support; on Baidu Cloud from 2026-07-03
May 2026Video modelmonth only
NAVA
A Baidu research model that generates synchronised video and stereo audio, including speech, from a single prompt. Optional reference images and speaker audio give per-speaker voice control in multi-speaker scenes. It is a 6.3B-parameter model, and the weights are published under Apache 2.0.
- 6.3B parameters, Apache 2.0
- Joint video + stereo audio (scene and speech)
- 1280x704 output; ~6-10 second clips
- Multi-speaker voice control via reference audio
AccessOpen weights on Hugging Face (Apache 2.0)
9 May 2026Frontier LLM
ERNIE 5.1
Baidu's current flagship and a more efficient successor to ERNIE 5.0, with about one-third of the total parameters and half the active ones. Baidu highlights gains in search, agent tasks and creative writing, and reports it placing first among Chinese models on LMArena's search leaderboard.
- ~1/3 total, ~1/2 active parameters of ERNIE 5.0
- AIME26 (with tools) 99.6 (lab-reported)
- LMArena Search 1,223, 4th globally (lab-reported)
- Beats DeepSeek-V4-Pro on tau3-bench, SpreadsheetBench (lab)
AccessERNIE website/app (ernie.baidu.com, yiyan.baidu.com); Baidu AI Studio playground
15 Apr 2026Image model
ERNIE-Image
Baidu's open text-to-image model: an 8B diffusion transformer that is strong at rendering text for posters, infographics and UI-like images in English and Chinese. A Turbo version produces images in 8 steps instead of 50. Both are released under Apache 2.0 on Hugging Face.
- 8B diffusion transformer, Apache 2.0
- Dense text and layout rendering, English/Chinese
- Turbo variant: 8 inference steps
- GenEval 0.8856 with prompt enhancer (lab)
AccessOpen weights on Hugging Face (Apache 2.0), incl. ERNIE-Image-Turbo
29 Jan 2026Embedding/other
PaddleOCR-VL-1.5
A compact 0.9B vision-language model for document parsing from Baidu's PaddlePaddle team. It handles skewed or warped pages, seals, tables, formulas and charts, and merges tables across pages. Baidu reports state-of-the-art results on OmniDocBench v1.5 and its real-world photo variants.
- 0.9B parameters
- OmniDocBench v1.5: 94.5% (lab-reported)
- Tables, formulas, charts, seals; cross-page merging
- Multilingual, incl. Tibetan and Bengali
AccessPaddleOCR open-source release (weights location and licence not opened here)
22 Jan 2026Frontier LLM
ERNIE 5.0
Baidu's natively omni-modal foundation model: 2.4 trillion parameters in an ultra-sparse mixture-of-experts design, understanding and generating text, images, audio and video. It was previewed at Baidu World in November 2025, and the official version launched on 22 January 2026 at Baidu's 'Wenxin Moment' event.
- 2.4T total parameters, ~3% active per token
- Text, image, audio and video in and out
- Top Chinese model on LMArena text, Jan 2026 (lab blog)
- Preview first shown 2025-11-13 at Baidu World
AccessERNIE app/website (Wenxin Yiyan); Baidu cloud
Nov 2025Open-weights LLMmonth only
ERNIE-4.5-VL-28B-A3B-Thinking
An open vision-language reasoning model from the ERNIE 4.5 family that 'thinks with images', zooming into details as it reasons. Baidu claims it beats larger commercial models on several multimodal benchmarks while running on a single GPU. It is licensed under Apache 2.0.
- ~28-30B total / 3B active parameters (MoE)
- Visual reasoning: charts, STEM, grounding
- Image zoom and image search as tools
- Video understanding with event localisation
AccessOpen weights on Hugging Face (Apache 2.0)