Baidu Releases Ernie 4.5 Foundation Model and Ernie X1 Reasoning Model With Multimodal Capabilities

Baidu launched two new synthetic intelligence (AI) fashions on Sunday. First, the Ernie 4.5 is a basis mannequin that succeeds the corporate’s earlier era, whereas the second, Ernie X1, is a reasoning-focused mannequin. The latter can also be the Chinese tech large’s first try at a “deep-thinking” AI mannequin, and Baidu claims that the massive language mannequin (LLM) performs on-par with the DeepSeek-R1. Alongside the mannequin releases, Baidu additionally made its AI platform Ernie Bot free to entry, and customers won’t need to pay a premium to make use of the chatbot.

Baidu Releases New Ernie AI Models

In a press launch, the Chinese tech large introduced the discharge of the 2 new AI fashions. Currently, solely the Ernie 4.5 mannequin has been launched and is accessible by way of the chatbot platform, and the corporate mentioned that the Ernie X1 will likely be accessible quickly. Typically, AI corporations introduce new fashions with a analysis paper and repository (in the event that they’re open-source), nonetheless, Baidu didn’t launch both.

The firm highlighted that the Ernie 4.5 will likely be open-sourced on June 30 with supply code and mannequin weights made accessible to builders. No such announcement has been made in regards to the reasoning mannequin.

Ernie 4.5 is a natively multimodal AI mannequin with the power to know textual content, pictures, audio, and movies. This means customers can add these file codecs as enter and ask the AI queries about them. The firm additionally mentioned that the language capabilities of the mannequin are being enhanced and it could possibly now higher reply to conversational messages in addition to reasoning and memory-based queries. Ernie 4.5 may also perceive contextual content material corresponding to memes, satire, and others.

Ernie 4.5 benchmark scores
Photo Credit: Baidu

As for efficiency, Baidu claimed that its newest basis mannequin outperforms GPT-4.5 throughout numerous benchmarks, and is priced at one p.c of GPT-4.5. The efficiency metrics are based mostly on the corporate’s inside testing. Apart from the AI platform, Ernie 4.5 can also be accessible by way of utility programming interfaces on Baidu AI Cloud’s MaaS platform Quianfan.

On pricing, Baidu mentioned that Ernie 4.5 enter costs begin at CNY 0.004 (roughly Rs. 0.04) per thousand tokens, and the output costs are set at CNY 0.016 (roughly Rs. 0.19) per thousand tokens.

Coming to Ernie X1, the corporate claimed that the multimodal reasoning mannequin comes with the aptitude of device use and excels in parameters corresponding to Chinese information Q&A, literary creation, manuscript writing, dialogue, logical reasoning, and advanced arithmetic. Some of the instruments it helps embrace superior search, picture understanding, doc understanding, picture era, code interpretation, webpage studying, TreeMind mapping, and extra.

The firm mentioned that its first reasoning mannequin was skilled utilizing progressive reinforcement studying (RL), an end-to-end coaching strategy, and a unified multi-faceted reward system. Just like DeepSeek-R1, it additionally comes with a visual chain of thought (CoT).

While Ernie X1 just isn’t accessible on the Quianfan Platform, as soon as it’s launched, its enter costs will likely be stored at CNY 0.002 (roughly Rs. 0.02) and output costs at CNY 0.008 (roughly Rs. 0.09) per thousand tokens, respectively.