Baidu releases open source model family ERNIE 4.5
Key Points
- Baidu has open-sourced its new ERNIE 4.5 model family under the Apache 2.0 license, reversing its previous closed-source approach and aiming to let developers build AI applications without high costs or vendor lock-in.
- The ERNIE 4.5 lineup includes ten variants, from small 0.3 billion parameter models to large Mixture-of-Experts versions with up to 424 billion total parameters, along with developer tools like ERNIEKit and FastDeploy.
- While ERNIE 4.5 outperforms other Chinese open-source models on most benchmarks and signals a global shift toward open AI development, it does not currently lead established models from OpenAI, Google, or Deepseek in overall performance.
Baidu is shifting its strategy in the global AI race by open-sourcing its new ERNIE 4.5 model family under the Apache 2.0 license. This marks a major reversal for the company, which previously kept ERNIE closed-source.
The move comes after Deepseek, a Chinese open-source model, drew international attention and put pressure on Baidu to respond. Baidu was one of the first companies worldwide to launch a ChatGPT competitor with its original Ernie, but the release of ERNIE 4.5 signals a clear pivot toward open development. According to CEO Robin Li, the goal is to enable developers around the world to build powerful AI applications without high costs or proprietary lock-in.
Strong benchmark results, but not leading the pack
ERNIE 4.5 includes ten variants, ranging from Mixture-of-Experts (MoE) models with up to 47 active and 424 billion total parameters to a compact 0.3 billion parameter version. Baidu has also released developer tools like ERNIEKit and the FastDeploy deployment toolkit.
The ERNIE-4.5-300B-A47B-Base model outperforms Deepseek-V3-671B-A37B-Base on 22 out of 28 benchmarks. There are no published comparisons with industry leaders like Deepseek's R1, OpenAI's o3, Claude 4, or Gemini 2.5 Pro. Even so, available results show ERNIE 4.5 can match or even surpass established models in several areas, though it doesn't take the top spot overall. Baidu has also developed X1, a direct competitor to R1, but it is not part of the current release.
Industry analysts see the ERNIE 4.5 release as a significant signal on the global stage. "Every time a major lab open-sources a powerful model, it raises the bar for the entire industry," Sean Ren of the University of Southern California told CNBC. He sees this as a challenge to companies like OpenAI and Anthropic, which rely on closed APIs and premium pricing. OpenAI CEO Sam Altman addressed the shift after Deepseek's R1 release, announcing that an open-source release is planned for the future.
More information and the models themselves are available on the ERNIE GitHub page.
AI News Without the Hype – Curated by Humans
As a THE DECODER subscriber, you get ad-free reading, our weekly AI newsletter, the exclusive "AI Radar" Frontier Report 6× per year, access to comments, and our complete archive.
Subscribe now