Connect with us

Jobs & Careers

Baidu’s ERNIE 4.5 is Built On a ‘Heterogeneous MoE’ Architecture

Published

on


Chinese tech giant Baidu has open-sourced the ERNIE 4.5 family of models on Monday, with numerous variants from 0.3B (billion) to 424B parameters. Baidu has released both language and multimodal (image and video) AI models under the ERNIE 4.5 umbrella, and all of them are publicly accessible under the Apache 2.0 license. 

The 300B variant of ERNIE-4.5 outperforms DeepSeek-V3 671B on several benchmarks, across general reasoning, math, and coding tasks. 

Besides, the 21B variant outperforms Alibaba’s Qwen3-30B-A3B on several math and reasoning benchmarks, despite containing 30% fewer parameters. 

The core architectural innovation of the ERNIE 4.5 models is their heterogeneous modality Mixture of Experts (MoE) structure. This was designed from the ground up for multimodal learning. It uses separate experts for text and vision, along with shared experts to integrate knowledge across all modalities. 

“These architectural choices ensure that both modalities are effectively represented, allowing for mutual reinforcement during training,” Baidu stated in the technical report. 

This technique also improves the computational efficiency. Since visual experts have one-third the intermediate dimension size of text experts, it reduces the computation for visual tokens by about 66%. 

Moreover, in text-only scenarios, vision experts can be skipped to reduce memory overhead, the authors noted. 

Source: x.com/rohanpaul_ai

Furthermore, the company stated that through a series of ‘extreme optimisations’, the largest model in the family achieved 47% Model FLOPs Utilisation (MFU) on NVIDIA H800 GPUs. Based on the training approach, the largest ERNIE 4.5 model was able to achieve “optimal training performance” with limited compute resources—around 96 GPUs, according to the company. 

Baidu stated that utilising techniques in pre-training, such as ‘intra-node’ expert parallelism, memory-efficient pipeline scheduling, FP8 mixed-precision training, and fine-grained recomputation methods, has helped achieve the stated levels of efficiency. 

For more details about the model’s architecture, pre-training and post-training techniques, and evaluations, refer to the PDF of the technical report.  



Source link

Jobs & Careers

NVIDIA Reveals Two Customers Accounted for 39% of Quarterly Revenue

Published

on



NVIDIA disclosed on August 28, 2025, that two unnamed customers contributed 39% of its revenue in the July quarter, raising questions about the chipmaker’s dependence on a small group of clients.

The company posted record quarterly revenue of $46.7 billion, up 56% from a year ago, driven by insatiable demand for its data centre products.

In a filing with the U.S. Securities and Exchange Commission (SEC), NVIDIA said “Customer A” accounted for 23% of total revenue and “Customer B” for 16%. A year earlier, its top two customers made up 14% and 11% of revenue.

The concentration highlights the role of large buyers, many of whom are cloud service providers. “Large cloud service providers made up about 50% of the company’s data center revenue,” NVIDIA chief financial officer Colette Kress said on Wednesday. Data center sales represented 88% of NVIDIA’s overall revenue in the second quarter.

“We have experienced periods where we receive a significant amount of our revenue from a limited number of customers, and this trend may continue,” the company wrote in the filing.

One of the customers could possibly be Saudi Arabia’s AI firm Humain, which is building two data centers in Riyadh and Dammam, slated to open in early 2026. The company has secured approval to import 18,000 NVIDIA AI chips.

The second customer could be OpenAI or one of the major cloud providers — Microsoft, AWS, Google Cloud, or Oracle. Another possibility is xAI.

Previously, Elon Musk said xAI has 230,000 GPUs, including 30,000 GB200s, operational for training its Grok model in a supercluster called Colossus 1. Inference is handled by external cloud providers. 

Musk added that Colossus 2, which will host an additional 550,000 GB200 and GB300 GPUs, will begin going online in the coming weeks. “As Jensen Huang has stated, xAI is unmatched in speed. It’s not even close,” Musk wrote in a post on X.Meanwhile, OpenAI is preparing for a major expansion. Chief Financial Officer Sarah Friar said the company plans to invest in trillion-dollar-scale data centers to meet surging demand for AI computation.

The post NVIDIA Reveals Two Customers Accounted for 39% of Quarterly Revenue appeared first on Analytics India Magazine.



Source link

Continue Reading

Jobs & Careers

‘Reliance Intelligence’ is Here, In Partnership with Google and Meta 

Published

on



Reliance Industries chairman Mukesh Ambani has announced the launch of Reliance Intelligence, a new wholly owned subsidiary focused on artificial intelligence, marking what he described as the company’s “next transformation into a deep-tech enterprise.”

Addressing shareholders, Ambani said Reliance Intelligence had been conceived with four core missions—building gigawatt-scale AI-ready data centres powered by green energy, forging global partnerships to strengthen India’s AI ecosystem, delivering AI services for consumers and SMEs in critical sectors such as education, healthcare, and agriculture, and creating a home for world-class AI talent.

Work has already begun on gigawatt-scale AI data centres in Jamnagar, Ambani said, adding that they would be rolled out in phases in line with India’s growing needs. 

These facilities, powered by Reliance’s new energy ecosystem, will be purpose-built for AI training and inference at a national scale.

Ambani also announced a “deeper, holistic partnership” with Google, aimed at accelerating AI adoption across Reliance businesses. 

“We are marrying Reliance’s proven capability to build world-class assets and execute at India scale with Google’s leading cloud and AI technologies,” Ambani said.

Google CEO Sundar Pichai, in a recorded message, said the two companies would set up a new cloud region in Jamnagar dedicated to Reliance.

“It will bring world-class AI and compute from Google Cloud, powered by clean energy from Reliance and connected by Jio’s advanced network,” Pichai said. 

He added that Google Cloud would remain Reliance’s largest public cloud partner, supporting mission-critical workloads and co-developing advanced AI initiatives.

Ambani further unveiled a new AI-focused joint venture with Meta. 

He said the venture would combine Reliance’s domain expertise across industries with Meta’s open-source AI models and tools to deliver “sovereign, enterprise-ready AI for India.”

Meta founder and CEO Mark Zuckerberg, in his remarks, said the partnership is aimed to bring open-source AI to Indian businesses at scale. 

“With Reliance’s reach and scale, we can bring this to every corner of India. This venture will become a model for how AI, and one day superintelligence, can be delivered,” Zuckerberg said.

Ambani also highlighted Reliance’s investments in AI-powered robotics, particularly humanoid robotics, which he said could transform manufacturing, supply chains and healthcare. 

“Intelligent automation will create new industries, new jobs and new opportunities for India’s youth,” he told shareholders.

Calling AI an opportunity “as large, if not larger” than Reliance’s digital services push a decade ago, Ambani said Reliance Intelligence would work to deliver “AI everywhere and for every Indian.”

“We are building for the next decade with confidence and ambition,” he said, underscoring that the company’s partnerships, green infrastructure and India-first governance approach would be central to this strategy.

The post ‘Reliance Intelligence’ is Here, In Partnership with Google and Meta  appeared first on Analytics India Magazine.



Source link

Continue Reading

Jobs & Careers

Cognizant, Workfabric AI to Train 1,000 Context Engineers

Published

on


Cognizant has announced that it would deploy 1,000 context engineers over the next year to industrialise agentic AI across enterprises.

According to an official release, the company claimed that the move marks a “pivotal investment” in the emerging discipline of context engineering. 

As part of this initiative, Cognizant said it is partnering with Workfabric AI, the company building the context engine for enterprise AI. 

Cognizant’s context engineers will be powered by Workfabric AI’s ContextFabric platform, the statement said, adding that the platform transforms the organisational DNA of enterprises, how their teams work, including their workflows, data, rules, and processes, into actionable context for AI agents.Context engineering is essential to enabling AI a

Subscribe or log in to Continue Reading

Uncompromising innovation. Timeless influence. Your support powers the future of independent tech journalism.

Already have an account? Sign In.



Source link

Continue Reading

Trending