Latest news with #foundationModels
National Post
10-07-2025
- Business
- National Post
Liquid AI Releases World's Fastest and Best-Performing Open-Source Small Foundation Models
Article content Next-generation edge models outperform top global competitors; now available open source on Hugging Face Article content CAMBRIDGE, Mass. — Liquid AI announced today the launch of its next-generation Liquid Foundation Models (LFM2), which set new records in speed, energy efficiency, and quality in the edge model class. This release builds on Liquid AI's first-principles approach to model design. Unlike traditional transformer-based models, LFM2 is composed of structured, adaptive operators that allow for more efficient training, faster inference and better generalization – especially in long-context or resource-constrained scenarios. Article content Article content Liquid AI open-sourced its LFM2, introducing the novel architecture in full transparency to the world. LFM2's weights can now be downloaded from Hugging Face and are also available through the Liquid P layground for testing. Liquid AI also announced that the models will be integrated into its Edge AI platform and an iOS-native consumer app for testing in the following days. Article content 'At Liquid, we build best-in-class foundation models with quality, latency, and memory efficiency in mind,' said Ramin Hasani, co-founder and CEO of Liquid AI. 'LFM2 series of models is designed, developed, and optimized for on-device deployment on any processor, truly unlocking the applications of generative and agentic AI on the edge. LFM2 is the first in the series of powerful models we will be releasing in the coming months.' Article content The release of LFM2 marks a milestone in global AI competition and is the first time a U.S. company has publicly demonstrated clear efficiency and quality gains over China's leading open-source small language models, including those developed by Alibaba and ByteDance. Article content In head-to-head evaluations, LFM2 models outperform state-of-the-art competitors across speed, latency and instruction-following benchmarks. Key highlights: Article content LFM2 exhibits 200 percent higher throughput and lower latency compared to Qwen3, Gemma 3n Matformer and every other transformer- and non-transformer-based autoregressive models available to date, on CPU. The model not only is the fastest, but also on average performs significantly better than models in each size class on instruction-following and function calling (the main attributes of LLMs in building reliable AI agents). This places LFM2 as the ideal choice of models for local and edge use-cases. LFMs built based on this new architecture and the new training infrastructure show 300 percent improvement in training efficiency over the previous versions of LFMs, making them the most cost-efficient way to build capable general-purpose AI systems. Article content Shifting large generative models from distant clouds to lean, on‑device LLMs unlocks millisecond latency, offline resilience, and data‑sovereign privacy. These are capabilities essential for phones, laptops, cars, robots, wearables, satellites, and other endpoints that must reason in real time. Aggregating high‑growth verticals such as edge AI stack in consumer electronics, robotics, smart appliances, finance, e-commerce, and education, before counting defense, space, and cybersecurity allocations, pushes the TAM for compact, private foundation models toward the $1 trillion mark by 2035. Article content Liquid AI is engaged with a large number of Fortune 500 companies in these sectors. They offer ultra‑efficient small multimodal foundation models with a secure enterprise-grade deployment stack that turns every device into an AI device, locally. This gives Liquid AI the opportunity to obtain an outsized share on the market as enterprises pivot from cloud LLMs to cost-efficient, fast, private and on‑prem intelligence. Article content Article content Article content Article content Article content Article content
Tahawul Tech
03-07-2025
- Business
- Tahawul Tech
Meta Superintelligence Labs Archives
Zuckerberg explained in the memo MSL will host the company's foundation models, products, and Fundamental AI Research teams, 'as well as a new lab focused on developing the next generation of our models'.
TechCrunch
06-06-2025
- Business
- TechCrunch
The best ways to build on top of foundation models, with DeepMind, Twelve Labs, and Amazon
New AI models are dropping all the time, and existing models are getting better all the time. How can a startup build a sustainable business on top of all of changing infrastructure and evolving foundation models? At TechCrunch Sessions: AI event, industry experts from DeepMind, Twelve Labs, and Amazon all shared their strategies for growing with foundation models instead of being left behind. In the minds of these experts, that challenge also presents an opportunity: Founders get the chance to build what they really want to build, with more tools and resources out the gate than ever before, thanks to existing AI infrastructure that's only improving.
Coin Geek
30-05-2025
- Business
- Coin Geek
Could foundation models make RAG obsolete?
Homepage > News > Tech > Could foundation models make RAG obsolete? Getting your Trinity Audio player ready... This post is a guest contribution by George Siosi Samuels , managing director at Faiā. See how Faiā is committed to staying at the forefront of technological advancements here . Even the smartest systems can become outdated if the paradigm shifts. Reid Hoffman recently argued that it's not the end of RAG—Retrieval Augmented Generation. But for those of us watching the evolution of large language models (LLMs) through a sharper lens, the writing might already be on the wall. Just as Yahoo's exhaustive web directory model was outpaced by Google's (NASDAQ: GOOGL) probabilistic search engine, RAG may soon find itself outdated in the face of increasingly powerful foundation models. It's not about whether RAG works. It's about whether it will matter. From Yahoo to Google: A signal from the past To understand the trajectory we're on, we need only look back. Yahoo believed in curating the Internet. Directories. Taxonomies. Human-reviewed indexes. But Google introduced a radically different idea: don't catalog everything—just rank relevance dynamically. Instead of organizing knowledge beforehand, Google inferred what mattered most through algorithms and backlinks. That wasn't just a technological improvement—it was a shift in philosophy. A move from structure to signal. From effortful storage to elegant retrieval. RAG, in many ways, feels like Yahoo. It's a bolted-on system that tries to enhance LLMs by grafting in 'clean,' retrievable knowledge from databases and vector stores. The goal is noble: improve the factuality and trustworthiness of artificial intelligence (AI) responses by injecting it with curated context. But what if that need disappears? Why RAG feels like a transitional technology RAG solves a real problem: hallucination. LLMs, especially in their earlier versions, had a tendency to fabricate facts. By adding a retrieval layer—pulling in external documents to ground the generation—RAG helped bridge the gap between generative flexibility and factual precision. But in solving one problem, it introduces others: Latency and complexity : RAG pipelines require orchestration between multiple components—vector databases, embedding models, retrievers, and re-rankers. : RAG pipelines require orchestration between multiple components—vector databases, embedding models, retrievers, and re-rankers. Data management burden : Enterprises must constantly update and maintain high-quality corpora, often requiring labor-intensive cleanup and formatting. : Enterprises must constantly update and maintain high-quality corpora, often requiring labor-intensive cleanup and formatting. Hard to generalize: RAG systems perform well in narrow domains but can break or return noise when facing edge cases or unfamiliar queries. It feels like scaffolding. Useful during construction—but not part of the finished architecture. Inference is eating Search Recent breakthroughs in LLM capabilities suggest that we're entering a new paradigm—one where inference can increasingly replace retrieval. With the emergence of models like GPT-4o, Claude 3 Opus, and even Google Gemini Pro 2.5, we're witnessing: Longer context windows : These models can now ingest and reason over hundreds of pages of content without needing external retrieval mechanisms. : These models can now ingest and reason over hundreds of pages of content without needing external retrieval mechanisms. Better zero-shot performance : The models are learning to generalize across vast domains without needing hand-fed examples or fine-tuned prompts. : The models are learning to generalize across vast domains without needing hand-fed examples or fine-tuned prompts. Higher factual accuracy: As foundation models train on more comprehensive data, their inherent 'memory' becomes more useful than brittle plug-ins or patched-on sources. In other words, the model itself is the database. This mirrors Google's dominance over Yahoo. When Google proved you didn't need to manually catalog the Internet to find useful content, the race was over. In the same way, when LLMs can consistently generate accurate answers without needing retrieval scaffolding, the RAG era ends. Enterprise blockchain implications So why does this matter to the blockchain and Web3 space? Because the architecture of how we store and surface data is changing. In the past, enterprise blockchain projects focused heavily on data provenance , auditability , and structured information flows . Naturally, RAG-like systems seemed appealing—pair a blockchain ledger (structured, secure) with a retriever that could pull trusted data into AI responses. But if inference can outpace retrieval—if models become so strong that they infer trustworthiness based on deep pretraining and internal reasoning—the value of these data layer bolt-ons will shift. It could go three ways: Legacy enterprise solutions double down on RAG-like hybrids, bolting AI onto databases and chains for compliance reasons. Next-gen startups skip RAG entirely, trusting LLMs' inference power and layering blockchain only for verifiability , not retrieval. A new form of 'self-attesting' data emerges, where models generate and verify their own responses using on-chain signals—but without traditional RAG scaffolding. Blockchain, in this context, becomes a reference point , not a library. The foundation model becomes both the interface and the reasoner. Is clean data still necessary? One of the assumptions keeping RAG alive is this: clean data = better output. That's partially true. But it's also a bit of an old-world assumption. Think about Gmail, Google Drive, or even Google Photos. You don't have to organize these meticulously. You just type and Google finds . The same is starting to happen with LLMs. You no longer need perfectly labeled, indexed corpora. You just need volume and diverse context —and the model figures it out. Clean data helps, yes. However, the new AI paradigm values signal density more than signal purity . The cleaner your data, the less your model has to guess. But the better your model, the more it can intuit even from messy, unstructured information. That's a core shift—and one that should change how enterprises think about knowledge management and blockchain-based storage. RAG's final role: A stepping stone, not a standard So, where does this leave RAG? Likely as a valuable bridge—but not the destination. We'll probably still see RAG-like systems in regulated industries and legacy enterprise stacks for a while. But betting on the future of AI on retrieval is like betting on the future of maps on phonebooks. The terrain is changing. Foundation models won't need retrieval in the way we think about it today. Their training and inference engines will absorb and transmute information in ways that feel alien to traditional IT logic. Blockchain will still play a role—especially in authentication and timestamping—but less as a knowledge base, and more as a consensus layer that LLMs can reference like a cryptographic compass. Conclusion: The search for simplicity RAG helped patch early AI flaws. But patchwork can't match the architecture. The best technologies don't just solve problems—they disappear . Google didn't ask users to understand PageRank. It simply worked. In the same way, the most powerful LLMs won't require RAG—they'll simply respond with clarity and resonance. And that's the signal we should be tracking. In order for artificial intelligence (AI) to work right within the law and thrive in the face of growing challenges, it needs to integrate an enterprise blockchain system that ensures data input quality and ownership—allowing it to keep data safe while also guaranteeing the immutability of data. Check out CoinGeek's coverage on this emerging tech to learn more why Enterprise blockchain will be the backbone of AI. Watch | IEEE COINS Conference: Intersection of blockchain, AI, IoT & IPv6 technologies title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen="">
Associated Press
23-05-2025
- Business
- Associated Press
MBZUAI Launches Institute of Foundation Models and Establishes Silicon Valley AI Lab
SAN FRANCISCO, May 23, 2025 /PRNewswire/ -- Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has expanded its global footprint with the launch of its Institute of Foundation Models (IFM). The IFM is a multi-site initiative consisting of a newly established Silicon Valley Lab in Sunnyvale, CA, combined with previously announced lab facilities in Paris and Abu Dhabi. The launch event yesterday at the Computer History Museum in Mountain View, establishes the third node in its global research network. This strategic expansion connects the university with California's vibrant ecosystem of AI researchers, startups, and tech companies. For the UAE and MBZUAI, this move represents another strategic step in the country's long-term economic diversification plan. By investing in cutting-edge technologies like advanced AI foundation models, the UAE continues to build knowledge-based sectors to support its long-term economic and social transformation efforts. 'Today's launch of the IFM represents a major step forward for the collaboration and global development of frontier-class AI foundation models,' said Professor Eric Xing, President and University Professor, MBZUAI. 'Our expansion into Silicon Valley provides a critical footprint to grow our presence in one of the most vibrant AI ecosystems in the world. We're creating pathways for knowledge exchange with leading institutions and accessing a talent pool that understands how to scale research into real-world applications.' The launch event drew representatives from the world's leading AI companies and academic institutions, highlighting the growing interest in MBZUAI's global approach to foundation model research. At the heart of MBZUAI's demonstrations was PAN, a world model capable of infinite simulations of diverse realities ranging from basic physical interactions to complex agent scenarios. Unlike previous systems focused primarily on generating text, audio, or images, PAN predicts comprehensive world states by integrating multimodal inputs like language, video, spatial data, and physical actions. This enables advanced reasoning, strategic planning, and nuanced decision-making for applications from autonomous driving to robotics. PAN's innovative hierarchical architecture supports multi-level reasoning and real-time interaction within simulations, maintaining high accuracy over extended scenarios. Its companion, PAN-Agent, showcases its utility in multimodal reasoning tasks, such as mathematics and coding, within dynamic simulated environments. K2 and JAIS: Advanced Foundation Models with Global Impact The IFM lab is also advancing two flagship AI systems demonstrating the commitment to further advance frontier-class foundation models: K2 and JAIS. A soon to be released update to K2-65B will focus on delivering breakthrough reasoning capabilities with sustainable performance. JAIS stands as the world's most advanced Arabic large language model. At the IFM JAIS will continue to expand in capability with increased language support and add more context to preserve and promote the cultures it supports. Building AI in the Open: Transparency as a Core Value MBZUAI has established one of the industry's most transparent approaches to AI development, open-sourcing not just models but entire development processes—positioning IFM as a leader in building openly. The LLM360 initiative provides researchers with complete materials including training code, datasets, and model checkpoints. This openness is balanced with safeguards including international advisory boards and peer review processes that maintain research integrity. The IFM's structure includes dedicated teams focused on model architecture, training methods, evaluation frameworks, and safety systems—combining the agility of a startup with the resources of an established research institution. About Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) MBZUAI is a research-focused university in Abu Dhabi, and the first university dedicated entirely to the advancement of science through AI. For more information, visit For press inquiries: Aya Sakoury Head of PR and Strategic Communications [email protected] Photo: View original content to download multimedia: SOURCE MBZUAI
