Baidu unveils proprietary ERNIE 5 beating GPT-5 efficiency on charts, doc understanding and extra

Mere hours after OpenAI up to date its flagship basis mannequin GPT-5 to GPT-5.1, promising lowered token utilization general and a extra nice persona with extra preset choices, Chinese language search large Baidu unveiled its next-generation basis mannequin, ERNIE 5.0, alongside a collection of AI product upgrades and strategic worldwide expansions.

The objective: to place as a world contender within the more and more aggressive enterprise AI market.

Introduced on the firm's Baidu World 2025 occasion, ERNIE 5.0 is a proprietary, natively omni-modal mannequin designed to collectively course of and generate content material throughout textual content, photographs, audio, and video.

Not like Baidu’s lately launched ERNIE-4.5-VL-28B-A3B-Pondering, which is open supply beneath an enterprise-friendly and permissive Apache 2.0 license, ERNIE 5.0 is a proprietary mannequin and is accessible solely through Baidu’s ERNIE Bot web site (I wanted to pick it manuallyu from the mannequin picker dropdown) and the Qianfan cloud platform software programming interface (API) for enterprise clients.

Alongside the mannequin launch, Baidu launched main updates to its digital human platform, no-code instruments, and general-purpose AI brokers — all focused at increasing its AI footprint past China.

The corporate additionally launched ERNIE 5.0 Preview 1022, a variant optimized for text-intensive duties, alongside the overall preview mannequin that balances throughout modalities.

Baidu emphasised that ERNIE 5.0 represents a shift in how intelligence is deployed at scale, with CEO Robin Li stating: “While you internalize AI, it turns into a local functionality and transforms intelligence from a price right into a supply of productiveness.”

The place ERNIE 5.0 outshines GPT-5 and Gemini 2.5 Professional

ERNIE 5.0’s benchmark outcomes recommend that Baidu has achieved parity—or near-parity—with the highest Western basis fashions throughout a large spectrum of duties.

In public benchmark slides shared through the Baidu World 2025 occasion, ERNIE 5.0 Preview outperformed or matched OpenAI’s GPT-5-Excessive and Google’s Gemini 2.5 Professional in multimodal reasoning, doc understanding, and image-based QA, whereas additionally demonstrating robust language modeling and code execution talents.

The corporate emphasised its capacity to deal with joint inputs and outputs throughout modalities, slightly than counting on post-hoc modality fusion, which it framed as a technical differentiator.

On visible duties, ERNIE 5.0 achieved main scores on OCRBench, DocVQA, and ChartQA, three benchmarks that take a look at doc recognition, comprehension, and structured knowledge reasoning.

Baidu claims the mannequin beat each GPT-5-Excessive and Gemini 2.5 Professional on these doc and chart-based benchmarks, areas it describes as core to enterprise functions like automated doc processing and monetary evaluation.

In picture era, ERNIE 5.0 tied or exceeded Google’s Veo3 throughout classes together with semantic alignment and picture high quality, in keeping with Baidu’s inner GenEval-based analysis. Baidu claimed that the mannequin’s multimodal integration permits it to generate and interpret visible content material with better contextual consciousness than fashions counting on modality-specific encoders.

For audio and speech duties, ERNIE 5.0 demonstrated aggressive outcomes on MM-AU and TUT2017 audio understanding benchmarks, in addition to query answering from spoken language inputs. Its audio efficiency, whereas not as closely emphasised as imaginative and prescient or textual content, suggests a broad functionality footprint meant to assist full-spectrum multimodal functions.

In language duties, the mannequin confirmed robust outcomes on instruction following, factual query answering, and mathematical reasoning—core areas that outline the enterprise utility of enormous language fashions.

The Preview 1022 variant of ERNIE 5.0, tailor-made for textual efficiency, confirmed even stronger language-specific leads to early developer entry. Whereas Baidu doesn’t declare broad superiority generally language reasoning, its inner evaluations recommend that ERNIE 5.0 Preview 1022 closes the hole with top-tier English-language fashions and outperforms them in Chinese language-language efficiency.

Whereas Baidu didn’t launch full benchmark particulars or uncooked scores publicly, its efficiency positioning suggests a deliberate try to border ERNIE 5.0 not as a distinct segment multimodal system however as a flagship mannequin aggressive with the most important closed fashions in general-purpose reasoning.

The place Baidu claims a transparent lead is in structured doc understanding, visible chart reasoning, and integration of a number of modalities right into a single, native modeling structure. Impartial verification of those outcomes stays pending, however the breadth of claimed capabilities positions ERNIE 5.0 as a critical different within the multimodal basis mannequin panorama.

Enterprise Pricing Technique

ERNIE 5.0 is positioned on the premium finish of Baidu’s mannequin pricing construction. The corporate has launched particular pricing for API utilization on its Qianfan platform, aligning the price with different top-tier choices from Chinese language opponents like Alibaba.

The distinction in price between ERNIE 5.0 and earlier fashions corresponding to ERNIE 4.5 Turbo underscores Baidu’s technique to differentiate between high-volume, low-cost fashions and high-capability fashions designed for complicated duties and multimodal reasoning.

In comparison with different U.S. alternate options, it stays mid-range in pricing:

World Enlargement: Merchandise and Platforms

In tandem with the mannequin launch, Baidu is increasing internationally:

• GenFlow 3.0, now with 20M+ customers, is the corporate’s largest general-purpose AI agent and options enhanced reminiscence and multimodal process dealing with.

• Famou, a self-evolving agent able to dynamically fixing complicated issues, is now commercially accessible through invite.

• MeDo, the worldwide model of Baidu’s no-code builder Miaoda, is stay globally through medo.dev.

• Oreate, a productiveness workspace with doc, slide, picture, video, and podcast assist, has reached over 1.2M customers worldwide.

Baidu’s digital human platform, already rolled out in Brazil, can be a part of the worldwide push. In response to firm knowledge, 83% of livestreamers throughout this yr’s “Double 11” buying occasion in China used Baidu’s digital human tech, contributing to a 91% enhance in GMV.

In the meantime, Baidu’s autonomous ride-hailing service Apollo Go has surpassed 17 million rides, working driverless fleets in 22 cities and claiming the title of the world’s largest robotaxi community.

Open-Supply Imaginative and prescient-Language Mannequin Garners Business Consideration

Two days earlier than the flagship ERNIE 5.0 occasion, Baidu additionally launched an open-source multimodal mannequin beneath the Apache 2.0 license: ERNIE-4.5-VL-28B-A3B-Pondering.

As reported by my colleague Michael Nuñez at VentureBeat, the mannequin prompts simply 3 billion parameters whereas sustaining a complete of 28 billion, utilizing a Combination-of-Consultants (MoE) structure for environment friendly inference.

Key technical improvements embrace:

• “Pondering with Photos”, which permits dynamic zoom-based visible evaluation

• Assist for chart interpretation, doc understanding, visible grounding, and temporal consciousness in video

• Runtime on a single 80GB GPU, making it accessible to mid-sized organizations

• Full compatibility with Transformers, vLLM, and Baidu’s FastDeploy toolkits

This launch provides stress on closed-source opponents. With Apache 2.0 licensing, ERNIE-4.5-VL-28B-A3B-Pondering turns into a viable basis mannequin for business functions with out licensing restrictions — one thing few high-performing fashions on this class supply.

Group Suggestions and Baidu’s Response

Following the launch of ERNIE 5.0, developer and AI evaluator Lisan al Gaib (@scaling01) posted a blended assessment on X. Whereas initially impressed by the mannequin’s benchmark efficiency, they reported a persistent challenge the place ERNIE 5.0 would repeatedly invoke instruments — even when explicitly instructed to not — throughout SVG era duties.

“ERNIE 5.0 benchmarks seemed insane till I examined it… sadly it’s RL braindamaged or they’ve a critical challenge with their chat platform / system immediate,” Lisan wrote.

In a matter of hours, Baidu’s developer-focused assist account, @ErnieforDevs, responded:

“Thanks for the suggestions! It’s a identified bug — sure syntax can constantly set off it. We’re engaged on a repair. You’ll be able to strive rephrasing or altering the immediate to keep away from it for now.”

The fast turnaround displays Baidu’s growing emphasis on developer communication, particularly because it courts worldwide customers via each proprietary and open-source choices.

Outlook for Baidu and its ERNIE foundational LLM household

Baidu’s ERNIE 5.0 marks a strategic escalation within the international basis mannequin race. With efficiency claims that put it on par with probably the most superior methods from OpenAI and Google, and a mixture of premium pricing and open-access alternate options, Baidu is signaling its ambition to develop into not only a home AI chief, however a reputable international infrastructure supplier.

At a time when enterprise AI customers are more and more demanding multimodal efficiency, versatile licensing, and deployment effectivity, Baidu’s two-track method—premium hosted APIs and open-source releases—could broaden its attraction throughout each company and developer communities.

Whether or not the corporate’s efficiency claims maintain up beneath third-party testing stays to be seen. However in a panorama formed by rising prices, mannequin complexity, and compute bottlenecks, ERNIE 5.0 and its supporting ecosystem give Baidu a aggressive place within the subsequent wave of AI deployment.