Baidu was early to large models (大模型) with its ERNIE (文心) family, but for years kept the weights closed. In mid-2025 it changed course in a way few expected: it gave away an entire model family at once.
Key takeaways
- Baidu open-sourced ERNIE 4.5 on 30 June 2025, releasing ten checkpoints with weights and inference code under the Apache 2.0 licence.
- The family spans mixture-of-experts (MoE) models with 47B and 3B active parameters—the largest at 424B total—plus a 0.3B dense model, with a 128K-token context.
- ERNIE Bot (文心一言) had already been made free for all users on 1 April 2025, a month after the hosted ERNIE 4.5 and reasoning model X1 launched.
- A thinking variant, ERNIE-4.5-21B-A3B-Thinking, later topped Hugging Face's global trending chart.
- Baidu then unveiled ERNIE 5.0: a 2.4-trillion-parameter native omni-modal model covering text, image, video and audio in one framework, with under 3% of parameters active per token.
From closed flagship to open family
ERNIE is Baidu's knowledge-enhanced large-model line. The chatbot ERNIE Bot opened to the public in August 2023, and ERNIE 4.0 followed that October. For most of 2023 and 2024 the headline models stayed behind Baidu's cloud API. That changed on 30 June 2025, when Baidu published the full ERNIE 4.5 series as open weights.
The release was unusually broad. Rather than a single model, Baidu shipped ten checkpoints: text and vision-language MoE variants at several scales (47B and 3B activated; up to 424B total parameters) and a 0.3B dense model, each with base and post-trained versions. All were released under Apache 2.0 with weights and inference code, downloadable from Hugging Face, GitHub and Baidu's own PaddlePaddle (飞桨) community. Chinese state media noted that the move made Baidu the first major Chinese internet company to achieve "double-layer open source"—both the framework and the model.
Why ten sizes matters
Most open releases lead with one or two flagships. Baidu's ten-checkpoint spread is a deployment strategy. A 0.3B dense model runs on modest hardware; the 21B-A3B variant targets cost-sensitive servers; the 424B total model is for those who want the top tier. By open-sourcing the whole ladder at once, Baidu lets enterprises and researchers pick a size that fits their compute instead of being forced onto a single API price.
The models also carried a genuine architecture claim: a multimodal heterogeneous MoE that shares parameters across text and vision while keeping dedicated experts per modality, so adding vision does not erode text quality. Baidu reported a 47% model-FLOPs-utilisation on the largest language pretraining—a systems-efficiency number that matters more to operators than to benchmark watchers.
Toward omni-modal
Baidu did not stop at 4.5. It unveiled ERNIE 5.0 in late 2025 and updated it in January 2026: a 2.4-trillion-parameter "native omni-modal" model trained from scratch to handle text, image, video and audio inside a single autoregressive framework, with fewer than 3% of parameters activated per token. In January 2026, the ERNIE-5.0-0110 checkpoint scored 1,460 on the LMArena text leaderboard—first among Chinese models and eighth globally, by Baidu's published results.
The trajectory is clear: from a text-and-vision MoE you can self-host, Baidu is betting that the next frontier is a single model that generates across every modality, not a suite of specialised ones.
What Apache 2.0 changes for buyers
The licence choice matters more than the parameter counts. Apache 2.0 permits commercial use, modification and redistribution without a per-seat fee or a revenue threshold, which removes the main legal friction that keeps many enterprises on closed APIs. For banks, hospitals and government contractors in China, that opens a path to run ERNIE 4.5 entirely on their own infrastructure—on Baidu's PaddlePaddle stack or a compatible framework—rather than sending data to a cloud endpoint. Baidu reinforced this by open-sourcing the training and deployment toolkits alongside the weights, so the moat shifts from "only we can run this model" to "our tooling makes it the easiest to run."
The trade-off is bifurcation. The 4.5 family is open; the newer 5.0 omni-modal core, at least at launch, is positioned as Baidu's flagship and is not part of the Apache release. A developer who builds a product on open ERNIE 4.5 may find the cutting-edge capability locked behind Baidu's commercial API—a familiar pattern across the industry, but worth pricing into any long-term architecture decision.
Honest limitations
The parameter counts, licence terms and checkpoint list for ERNIE 4.5 come from Baidu's official blog and Chinese state-media coverage; we have not downloaded and verified every checkpoint's weight integrity. The 2.4-trillion-parameter and LMArena figures for ERNIE 5.0 are Baidu's own disclosures and are not independently audited. "Omni-modal" performance depends heavily on the evaluation suite; cross-modal generation quality is best judged by hands-on use, not leaderboard scores. The Apache 2.0 release covers the 4.5 family, not necessarily every future ERNIE variant.
What readers can do now
- If you need a self-hostable Chinese model with vision, start from the ERNIE 4.5 Apache 2.0 checkpoints and match the size to your GPU budget.
- For production, compare the 21B-A3B variant against Qwen and DeepSeek peers on your own tasks—MoE active-parameter counts, not total size, predict your cost.
- Watch whether Baidu open-weights ERNIE 5.0's omni-modal core; if it stays closed, the open line and the flagship will diverge.