Baidu AI Day 2024: Wenxin X1 Turbo Sets New Benchmark with Top‑Level Evaluation and Advanced Multimodal Capabilities

On May 20, Baidu held its AI Day in Beijing where senior executives, including Vice President Wu Tian, presented the latest advancements of the Wenxin large‑model series, emphasizing the new 4.5 Turbo version and the upgraded X1 Turbo model with superior multimodal, reasoning, and tool‑calling abilities.

The 4.5 Turbo model introduces mixed training of text, images, and video using heterogeneous expert modeling, adaptive resolution visual encoding, three‑dimensional rotational positional encoding, and modality‑aware loss calculations, boosting cross‑modal learning efficiency by nearly twofold and improving understanding performance by over 30%.

In post‑training, Baidu deployed a self‑feedback enhancement framework that creates a "train‑generate‑feedback‑enhance" loop, reducing data production costs, mitigating hallucinations, and markedly improving the model’s handling of complex tasks.

Training also incorporates preference‑based reinforcement learning with unified reward mechanisms, enhancing result quality assessment, data utilization, and stability, while simultaneously advancing the model’s comprehension, generation, logical reasoning, and memory capabilities.

Innovations in reasoning combine tool invocation with chain‑of‑thought prompting, forming a composite thinking‑action chain that yields clearer, more logical outputs and expands cross‑domain problem‑solving.

A closed‑loop data pipeline—"data mining & synthesis → data analysis & evaluation → model capability feedback"—ensures continuous production of high‑density, diverse, domain‑rich data, and is easily extensible to new data types.

Applications showcased include hyper‑realistic digital humans driven by script‑based multimodal collaboration, achieving over 100,000 AI anchors with a 31% conversion rate and an 80% reduction in live‑streaming costs, as well as the Wenxin Code Assistant (Wenxin KuaiMa), which now generates more than 40% of daily new code for developers, serving 7.6 million users.

The China Academy of Information and Communications Technology released its large‑model inference evaluation, granting Wenxin X1 Turbo a top‑level "4+" rating—16 categories scoring 5, 7 scoring 4, and 1 scoring 3—making it the first domestic model to pass this assessment.

Evaluators highlighted X1 Turbo’s strong structured logical reasoning, balanced efficiency, robust data mechanisms, and enhanced safety, positioning it for broad application across industries.

Student user Chen Junhang, a 16‑year‑old high‑schooler, shared how Wenxin’s conversational AI has become his constant study companion, likening it to an always‑available dictionary, and how he leveraged it to create a smart copy‑generation tool that dramatically reduced his mother’s workload.

The event underscored Baidu’s commitment to advancing AI technology, fostering ecosystem growth, and delivering societal benefits through accessible, high‑performance large‑model solutions.