The company whose founder once called open-source models "IQ tax" just gave its flagship model away.
Baidu's ERNIE (文心) was the face of China's closed-source bet. In 2026 it became open — weights, code, and a license that lets you ship it commercially.
What shipped, and when
On 30 June 2026, Baidu open-sourced the ERNIE 4.5 (文心 4.5) series. The release covers 10 models spanning dense and mixture-of-experts (MoE) variants, from a tiny 0.3B dense model up to a 300B MoE, with vision-language (VL) variants in between.
- Released under Apache 2.0 — free commercial use, modification, and redistribution.
- Fully open: pre-trained weights and inference code, published on Baidu's AI Studio hub and on Hugging Face.
- Paired with ERNIEKit (fine-tuning and training) and FastDeploy (inference acceleration), so the release is a toolkit, not just a file dump.
Baidu had telegraphed the move: in late February 2026 it said ERNIE 4.5 would launch on 16 March and open-source on 30 June. The June date held.
The reversal
This is a genuine strategic flip, not a rebrand. Baidu founder Robin Li (李彦宏) had publicly argued that open-source models were "智商税" — a tax on intelligence, implying they were economically foolish. The success of DeepSeek's open weights forced a rethink across the Chinese majors.
- At Baidu's 2026 Create conference, the stated focus shifted from "the model race" to "agent application landing" — shipping usable AI products rather than chasing benchmark supremacy.
- Earlier in the year, on 29 January 2026, Baidu also open-sourced PaddleOCR-VL-1.5, the first OCR model with "irregular-shape localization" (异形框定位) for skewed, folded, or curled documents. That detail matters for real-world scanning, not demos.
- Baidu's AI chip arm, Kunlunxin (昆仑芯), lit up a 30,000-card fully self-developed cluster able to train multiple hundred-billion-parameter models at once — the compute base that makes an open-model strategy credible.
Why it matters
Baidu now sits in the open-weight camp alongside Alibaba (Qwen / Tongyi, fully open), DeepSeek, Zhipu AI, and Moonshot. The tactical logic:
- Apache 2.0 plus 0.3B–7B edge models lets small and mid-sized enterprises run ERNIE on their own servers — a direct answer to OpenAI and Anthropic API pricing and to data-sovereignty worries inside China.
- The 300B flagship targets cloud and API-wrapper businesses that want a domestic alternative to closed US models.
The risk Baidu runs is monetization. Open weights mean less API lock-in; Baidu's answer is the Qianfan (千帆) cloud platform and enterprise services layered on top of the free models. Whether "open to pull developers, cloud to capture value" actually converts into cloud revenue is the open question.
The open-model wave Baidu joined
Baidu is late to a party Alibaba started. Alibaba's Qwen (Tongyi) went fully open early and used open weights to pull developers toward Alibaba Cloud — the "open to monetize the cloud" playbook Baidu is now copying. DeepSeek proved a smaller lab could set the agenda with open releases, and Zhipu and Moonshot followed. By mid-2026, open weights had become the Chinese default for foundation models, while US frontier labs (OpenAI, Anthropic, Google) stayed largely closed.
There is a regulatory dimension too. China has treated open models as a tool of "technical sovereignty" — domestic alternatives to closed foreign systems — which gives open releases a policy tailwind that US labs do not enjoy. That makes Baidu's flip less a change of heart and more a recognition of where the domestic center of gravity already sat.
The unresolved tension is governance. Fully open weights raise questions about safe deployment and misuse that closed APIs can police more easily. Baidu's bet is that developer adoption and a managed cloud path outweigh those risks — a calculation every open-model lab in China is now making.
What readers can do now
- If you deploy models, benchmark ERNIE 4.5's 0.3B–7B edge variants for on-premise or edge use; Apache 2.0 removes the licensing risk that surrounds many "open-but-conditional" releases.
- If you are a China enterprise, weigh Baidu Qianfan against Alibaba Qwen and DeepSeek for sovereign deployment — the differentiator is now tooling and support, not raw model access.
- If you watch the space, track whether Baidu's "open + cloud" play converts free adoption into measurable cloud revenue, rather than just developer goodwill.
Honest limitations
- The 30 June 2026 ERNIE 4.5 open-source release, the 10-model lineup, the Apache 2.0 license, and the weights-plus-code availability are confirmed on Baidu's AI Studio model hub and Hugging Face, and corroborated by Chinese tech press (aitanjin.ai) and Baidu's February 2026 guidance.
- We did not find an international wire (Reuters/Bloomberg) directly confirming the open-source release at writing; primary confirmation is Baidu's own distribution platforms plus Chinese tech coverage, so treat as well-sourced but not wire-confirmed.
- PaddleOCR-VL-1.5 (29 January 2026) and the 30,000-card Kunlunxin cluster are from a Changsha Kaifu district government notice (kaifu.gov.cn), an official source.
- Robin Li's "IQ tax" remark and the Create-conference strategy shift are widely reported Chinese tech-media context; we could not locate the primary video clip and present them as background, not verified quotation.
- No currency amounts are stated in this piece, so no RMB/USD/HKD conversion was applied.