Baidu open sources ERNIE 4.5: a new AI power play
Hello readers,
Baidu has taken a surprising turn in the global AI race.
After years of guarding its ERNIE model family, the company has open-sourced ERNIE 4.5 under the Apache 2.0 license. This is a meaningful move, not just because of what’s inside the models, but because of what it signals: a shift in mindset from proprietary dominance to open development.
What triggered this change? Many are pointing to Deepseek’s recent rise. When Deepseek released its powerful open-source models, it sparked a conversation across the AI world. Baidu, once a leader with its closed ERNIE Bot, now seems to be playing catch-up with a more open approach.
CEO Robin Li put it plainly:
“The goal is to let developers build powerful AI without high costs or vendor lock-in.”
What’s in ERNIE 4.5?
The release includes a broad model family:
- Ten versions, including Mixture-of-Experts (MoE) architectures up to 424B total parameters
- A compact 0.3B model for lightweight tasks
- ERNIEKit for development
- FastDeploy for scalable deployment
ERNIE-4.5-300B-A47B-Base outperformed Deepseek-V3-671B-A37B-Base on 22 out of 28 benchmarks, which is strong, but Baidu hasn’t yet compared it to models like Claude 4 or OpenAI’s o3. Its unreleased X1 model may be positioned for that.
Rysysth Insights
At Rysysth, we see this as less about model architecture and more about timing and signal.
Open-sourcing a flagship model in today’s climate is not just a technical release — it’s a message. It reflects growing peer pressure among AI labs to show openness, and it hints at how important global perception has become in this space.
What stands out to us is how fast things are shifting. Just six months ago, Chinese labs were being criticized for opacity. Now, they’re setting the pace in open development, while Western companies are under pressure to follow.
ERNIE 4.5 may not be the strongest model overall, but it doesn’t need to be. Its release is part of a larger narrative: AI dominance is no longer just about performance. It’s about access, transparency, and ecosystem influence.
That’s the part we’re watching closely. This isn’t just Baidu trying to stay relevant — it’s Baidu challenging others to rethink their playbooks.
In summary
- Baidu open-sourced ERNIE 4.5 after Deepseek’s growing influence
- The model family includes powerful MoE versions and dev toolkits
- Benchmarks are competitive, though not yet leading
- The move signals a deeper shift toward openness in global AI strategy
- It marks a cultural pivot for Chinese labs in how they engage the broader AI community
Until next time.