On February 14th, a groundbreaking announcement from Baidu erupted across the tech landscape, sending ripples of excitement and apprehension alike: the Wensi large model 4.5 series is set to launch and will be fully open-sourced starting June 30thJust a day earlier, Baidu had revealed another stunning move — Wensi Yiyan, their AI chatbot, would be completely free for all users, both on PC and mobile applications, alongside the rollout of enhanced deep search functionalities.

This series of announcements felt like a massive boulder dropped into a calm tech pond, creating waves that prompted many to wonder what this frenzy in the large model sector signifiesOpening the gates to open-source models is fast becoming a trend among industry giants, with heavyweights such as OpenAI and Baidu making significant advancements in rapid succession, dizzying observers with their pace and scale of innovation.

OpenAI acted promptly on February 6th, when it made its ChatGPT Search feature openly accessible, allowing users to dive right in without any barriersFollowing this, on February 13th, CEO Sam Altman unveiled details about the upcoming GPT-4.5 and GPT-5 models, promising unlimited interactions through the free version of ChatGPT and encouraging innovation without restrictionsMeanwhile, Baidu was matching this momentum with its own aggressive strategy — free access to Wensi Yiyan and the open-source launch of the Wensi large model series seemed to signal an intense competitive atmosphere.

The question lingering in the air was why open-sourcing had suddenly become so desirableBaidu's founder, Li Yanhong, offered insight during the World Governments Summit held on February 11th: "Open-source makes you more attractive." He elaborated that the realms of AI and generative AI are still in their nascent stages of innovationRapid dissemination of these technologies can accelerate adoption, inviting more players to explore and contribute, thus creating a promising cycle of growth and evolution.

In simpler terms, opening up models for public use can turbocharge their development and deployment

Advertisements

In the early stages of technology, the faster you can share and iterate, the quicker you can capture market attention and gather innovative forces, creating a snowball effect that propels the industry forward.

This leads to the intriguing question: why did OpenAI and Baidu decide on this particular timing for their open-source strategies? The truth lies in understanding that moving between open-source and proprietary approaches isn't a Black and White choice; it represents strategic considerations suited for different phases of a technology's lifecycle.

Initially, the notion of open-sourcing a model often came across more as a marketing tactic rather than a genuine plea for collaboration, demonstrated by efforts like Llama that offered only partial openness, leaving many feeling unsatisfiedYet, as we move into 2025, the large models have traversed from startup stages to a period of explosive applicationThis shift renders open-sourcing akin to launching a high-speed lane on an information superhighway, facilitating rapid knowledge transfer and presenting richer developmental opportunities for innovators.

In scrutinizing the strategies of OpenAI and Baidu, one quickly notes notable differences yet shared goalsOpenAI aims to open-source its "veteran" models that have already made an impact, while Baidu pushes forth with its latest offerings, the "new recruits." This contrast is akin to two shops — one auctioning off old goods while the other invites trial of novel products for free.

Baidu’s audacity with its open-source moves reflects a robust position, suggesting a commitment toward collaborative growth within the industryThe reality is that the open-source arena is for the seasoned and skilled; it isn't merely a playground for amateursWithout a firm grasp of technology, one risks being viewed as nothing but a fleeting curiosity in an increasingly competitive space.

However, amidst this open-source clamor, a critical aspect deserves to be spotlighted: beyond the buzz of open-source versus closed-source, what ultimately matters is the practical efficacy of the models themselves

Advertisements

Users primarily care about whether these models can successfully tackle complex tasks and the affordability of using these innovations.

With large models now entering a period of decreased costs and improved efficiencies, projections suggest that costs may reduce by over 90% annuallyAs these costs plummet, the tipping point for widespread application of large models draws tantalizingly closerLi Yanhong wisely articulated that "the critical factor isn't which model to use; whether it's open-source or closed is secondaryWhat counts is the creative value generated in practical applications."

For Chinese firms, this wave of open-sourcing during the AI era represents an unprecedented opportunity to shape a leading position on the global stageHistorically, the tech landscape has seen players like Windows/Linux and Apple/Android where local counterparts have often adapted rather than innovated.

However, the wind of open-source AI now offers an exceptional chance for a strategic leapfrogIt’s a common misconception that free open-source practices are only adopted by fledgling companiesIn contrast, Baidu is boldly embracing this approach, revealing a flurry of open policies in just three daysNews about Wensi 5.0 broke on February 12; the next day, Wensi Yiyan was announced as free; and by February 14, plans for next-generation model open-sourcing were unveiled.

Baidu's bravado in this arena could be likened to the "cave men" of the AI realm — a company armed with deep technological resources that acts without hesitation, plunging headfirst into the realms of free offerings and open-source modelsThis confidence springs from their dual advantage of lowered training and inference costs.

Contributions from their Kunlun chip have proven instrumental, presenting an exceptionally cost-effective solution that requires fewer computational resources while handling large-scale models like DeepSeek-V3/R1 with surprising efficiencyAs a result, the computational requirements for both inference and training have dropped significantly, consequently driving down associated costs.

Moreover, Baidu's intelligent cloud has successfully launched the Kunlun chip's third-generation Myriad cluster—this is the first domestically developed solution of its kind

Advertisements

Future enhancements aim for a capacity of 30,000 units, utilizing high-tech approaches such as task scheduling and elastic management to boost utilization rates dramatically.

Additionally, the Baijiao platform plays a role in resolving the significant bandwidth issues faced during large model training sessionsTheir sophisticated HPN high-performance network delivers improvements in congestion control and collective communication algorithms, attaining effectiveness levels exceeding 90%. Coupled with groundbreaking thermal management techniques that lower energy consumption, the overall costs for model training have notably reduced.

To ensure that large-scale clusters can reliably execute training tasks, the Baijiao platform also provides comprehensive diagnostic tools and accompanies them with proprietary BCCL technology, reducing downtime from hours to mere minutes, achieving an impressive training efficiency rate of 98%.

Turning our gaze toward inference costs, Baidu's four-layer technical architecture showcases its strengthsBy building an interlocking feedback loop across chips, frameworks, large models, and end-application scenarios, the model continuously optimizes itself over time, leading to reduced reasoning costsLi Yanhong released insights during an early 2024 earnings call, noting that Wensi's inference costs had already plunged to just 1% of what they were a year prior, marking an astonishing reduction.

For enterprises and developers working with the Wensi large model, benefiting from lower inference costs and higher training efficiencies unveils exciting possibilitiesSigns are emerging that Baidu will introduce multiple competitive large model options come 2025 — ranging from versions 4.5 to 5.0. Their investments and strategic placements in the AI landscape indicate a meticulously orchestrated marathonWith two generations of models released in 2023, a year solely focused on fostering applications in 2024, and all eyes set on a “model year” in 2025, Baidu's foundational strengths are positioned to elevate their ambitions to that of an industry leader.

Moreover, OpenAI's parallel releases of GPT-3.5 and 4.0 serve to illustrate a competitive synergy, with upcoming rollouts of 4.5 and 5.0 signaling that a head-to-head rivalry is brewing

The rapidly evolving AI sector is on the cusp of a significant application explosion, and Baidu has already begun its sprint towards the finish line.

Li Yanhong has remarked that the pace of innovation in large models outstrips traditional computing, with costs decreasing by as much as 90% every 12 months, affirming that "lowering costs is fundamentally intrinsic to innovation." The specter of a major application breakthrough driven by the reduction of expenses is no longer a distant horizonWith initiatives like DeepSeek lowering barriers to entry and Wensi being offered free, the eruption of large model applications is imminent.

The prior usage rates of Wensi large models demonstrated their dominance in the domestic sphere, and open-sourcing them is set to propel those figures to new heights while enlarging their scope of applicationBaidu's proactive and experimental attitude towards technology was witnessed as far back as 2016, when they launched the open-source PaddlePaddle framework, and their Qianfan large model platform remains one of the most integrated with models across the industry.

In this latest bout of large model competition, Baidu's technological heft and openness have secured an advantageous positionFurthermore, it's notable that Baidu has been at the forefront of large model industry implementations, emerging victorious in numerous benchmark projects in 2024. Even Apple has sought collaboration with Baidu to develop AI technologies for their iPhones in China, an endorsement of Baidu's technical prowess.

Despite the robust debates surrounding open-source versus proprietary models, the crux remains: can the applications truly deliver? No matter how powerful the models may be in performance and cost-effectiveness, lack of practical deployment reduces them to mere conceptsCurrently, as large models witness cost reductions, it feels akin to the dawn of a new age for application utilization, with Baidu positioned as a leader ready to catalyze this potential.

Baidu's Wensi large model boasts impressive capabilities, leveraging RAG and iRAG technologies as its competitive banners

Upon the launch of Wensi Yiyan, Baidu recognized the value inherent in enhanced retrieval, and over a year and a half, they have extracted substantial worth from this arena.

The synergy of Baidu's "understanding-retrieval-generation" technology allows the model to function like a savvy librarianIt comprehends inquiry before sourcing relevant text, subsequently synthesizing the information into precise and timely responses.

While this may sound simplistic, executing it effectively is quite the featAmong a plethora of domestic and international large models, Wensi Yiyan ranks decidedly high in this regardFor instance, when asked specific scenarios from popular media, such as “Why was Nezha uncomfortable while making lotus powder in Nezha 2?” or “What language did Huang Rong use to give Ouyang Feng the Nine Yin Manual in Tsui Hark's adaptation?”, it delivers the answers with clarity and depth, sometimes surpassing even ChatGPT's connected capabilities.

Baidu’s iRAG technology, is a noteworthy innovation in the realm of image generationBy harnessing the vast image library available through Baidu’s search, coupled with its formidable foundational model abilities, the realism of the produced images astounds, often devoid of the “AI” signatureAdditionally, the costs associated with image generation have plummeted dramatically.

iRAG boasts of its “four free” approach: free from hallucination, exceptionally realistic, economical, and ready for immediate useIf prompted to create an image of David Beckham, it will deliver precisely what you ask without producing something entirely unrelated.

This innovative technology finds application across diverse fields such as film, comic creation, storyboard illustration, and poster design while simultaneously addressing the common issues of AI-generated "hallucinations" and drastically reducing creative costsA typical vehicle advertisement project that would have traditionally cost tens of thousands is now nearly free with iRAG.

With such robust technology at its disposal, Baidu is strategically advancing in implementation

Advertisements

Advertisements