NASA & IBM Fuse 30 Data Layers into AI Moon Map, Unlocking Lunar Future

NASA is now using artificial intelligence to study the Moon in a whole new way. Working with IBM and top universities, they’ve built the NASA‑IBM Lunar Foundation Model — one of the first open‑source AI tools made just for lunar science. It’s trained on years of data from NASA’s Lunar Reconnaissance Orbiter and is freely available online. Anyone can explore it on Hugging Face or test the full code on GitHub, opening lunar research to the world.

NASA and IBM have released an open-source Lunar Foundation Model that fuses over 30 data layers from four missions into one AI map, improving lunar mapping accuracy by up to 23% and enabling better detection of ice, craters, and volcanic features. This breakthrough is directly tied to planning future lunar bases near the Moon’s south pole.

The NASA‑IBM Lunar Foundation Model transforms decades of Moon data into a discovery engine, revealing patterns no single mission could uncover.

A 10-image mosaic captured by NASA’s Lunar Reconnaissance Orbiter's Narrow Angle Camera between June 2012 and April 2016 showing the volcanic feature Mons Rümker and its surrounding mare plains. NASA/GSFC/Arizona State University

NASA’s AI ecosystem now spans Earth, Moon, and Sun: the Prithvi Model for geospatial Earth observation, the Lunar Foundation Model for Moon science, and the Surya Model for heliophysics. Together, they represent a “5+1” strategy to embed AI into every major science domain, enabling faster discoveries and operational applications.

Prithvi Model (Earth Observation)

• First orbital AI foundation model: Deployed aboard the ISS and South Australia’s Kanyini satellite in 2026.
• Training data: 13+ years of Harmonized Landsat & Sentinel‑2 imagery.
• Applications: Flood mapping, wildfire scar detection, crop yield prediction, land‑use monitoring.
• Strength: Performs analyses in orbit, reducing bandwidth needs by processing data before transmission.
• Open-source: Available on Hugging Face, enabling global researchers to fine‑tune for disaster response and environmental monitoring.

Lunar Foundation Model (Moon Science)

• Developed by NASA & IBM: Trained on 2M+ lunar image tiles from the Lunar Reconnaissance Orbiter.
• Capabilities:
  • Crater mapping for geological dating.
  • Ice prospectivity in permanently shadowed regions.
  • Volcanic feature detection (irregular mare patches).
  • Surface change detection (e.g., SpaceX rocket impact).
• Impact: Supports Artemis mission planning, safe landing site selection, and resource utilization for future lunar bases.

Surya Model (Heliophysics)

• Training data: 9 years of Solar Dynamics Observatory data.
• Architecture: Spatiotemporal transformer with spectral gating, designed for solar flare forecasting.
• Applications:
  • Forecasting solar flares up to 2 hours ahead.
  • Predicting solar wind speed and irradiance.
  • Tracking active regions and coronal mass ejections.
• Performance: Surpassed existing flare prediction benchmarks by 16%, offering early warnings for satellites, power grids, and aviation.
• Open-source: Hosted on Hugging Face with GitHub code, encouraging global collaboration.

Comparative Snapshot

Model	Domain	Training Data	Key Applications
Prithvi	Earth	Landsat + Sentinel‑2 (13 yrs)	Floods, crops, wildfires, land use
Lunar Foundation	Moon	LRO imagery (2M tiles)	Craters, ice, volcanic features
Surya	Sun	SDO (9 yrs, multi‑wavelength)	Solar flares, winds, irradiance

What the Model Does

• 30 data layers: Integrated from nine instruments across four missions, including the Lunar Reconnaissance Orbiter and GRAIL.
• Multimodal AI system: Combines thermal, topographic, gravitational, and multispectral data into one unified map.
• Accuracy gains: Ice prospectivity 22% reduction in error; crater detection 19% higher accuracy at 100m resolution using half the training data; volcanic mapping 3% improvement.

Built mainly on data from NASA’s Lunar Reconnaissance Orbiter, the model is freely available on Hugging Face, with its full codebase open on GitHub for anyone to test and experiment.

Why It Matters

• Lunar ice: Crucial for water, oxygen, and rocket fuel production. Permanently shadowed craters near the poles are prime targets.
• Safer landings: AI helps identify stable terrain for Artemis missions and future bases.
• Geological insights: Crater and volcanic mapping deepens understanding of the Moon’s history.
• Open-source availability: Released via Hugging Face and GitHub, enabling universities, startups, and smaller space agencies to build on it.

Key Comparisons

Feature	Improvement vs Baseline	Impact
Ice prospectivity	22% RMSE reduction	Better targeting of polar ice deposits
Crater detection	+19% accuracy, +19% mAP	Safer landing site selection
Volcanic mapping	+3% IoU	Improved geological history analysis
Training efficiency	Half the data needed	Accessible for smaller research teams

Challenges Ahead

• Domain shift: AI must adapt across different orbital sensors.
• Shadowed regions: Extremely cold, dark craters remain hard to study.
• Validation: Ground truth data from future missions is still needed.

Context for India

This development complements India’s Space Vision 2047, where AI-driven exploration is central to lunar resource utilization and base planning. It also parallels ISRO-backed projects like CraterMorpho, showing how global collaborations are converging on AI-powered planetary science.

NEC Collaborates with Japan’s Defense Innovation Institute on Next‑Gen Underwater Acoustic AI

• NEC Begins Research to Establish a Leading
• Underwater Acoustic Platform Model. 
• Awarded a Contract by Japan’s Agency for Defense Equipment,

Enabling High-Precision, Rapid Visualization of the Ocean Using Sound and AI~
NEC Corporation (NEC; TSE: 6701) has been awarded a contract for “Research on Underwater Acoustic Foundation Models Using Self-Supervised Learning (Part 1),” with a focus on dual-use applications, as part of the “Innovative Breakthrough Research” (*1) program conducted by the Defense Innovation Science and Technology Institute of Japan’s Agency for Defense Equipment. NEC has commenced research on this project and aims to complete development of a foundation model by fiscal year 2027.

This project involves training a model on vast amounts of underwater acoustic data to build a world-leading foundational model for underwater acoustics. The goal is to realize general-purpose generative AI, similar to large language models (LLMs), in the field of underwater acoustics.

Utilizing this foundational model will enable more accurate and rapid understanding of underwater phenomena and facilitate the creation of a digital twin of the ocean. This is expected to greatly enhance application possibilities, not only for defense purposes but also for areas such as the exploration of marine life, resources and environmental monitoring, as well as for high-precision earthquake prediction and other uses.

Overview of the Underwater Acoustics Foundation Model Under Development

Background

Oceans cover more than 70% of the Earth’s surface and hold many mysteries. Today, efforts are underway to promote their sustainable use—including the exploration of marine life, resources, and the prediction of events occurring on the seafloor. Since light and radio waves barely penetrate the water, the use of sound waves is an effective method for a wide range of observations and recognition tasks. Underwater sound waves are captured as underwater acoustic data using sonar and hydrophones, and are analyzed by skilled experts who employ advanced signal processing techniques and expertise. However, analyzing such vast amounts of data requires a great deal of time and concentration, and there are limitations to the level of accuracy that can be achieved.

About the Underwater Acoustic Foundation Model

While the development of language foundation models is advancing worldwide, few general-purpose foundation models for underwater acoustics have been realized, and development has been limited to AI models specialized for specific applications or tasks. Conventional AI models require large amounts of “labeled data,” making them difficult to apply to applications or tasks where only a small amount of labeled data can be collected. In the field of underwater acoustics in particular, collecting labeled data is challenging, making it difficult to build large-scale AI models.

Building on over 90 years of experience in sonar technology development, NEC will develop a general-purpose underwater acoustics foundation model. This effort will leverage the large-scale learning expertise gained through “cotomi” (*2), NEC’s proprietary AI technology, as well as AI technology based on self-supervised learning, which enables the utilization of large volumes of data even in the absence of labeled data. The large-scale underwater acoustics data used for training will be collected through public-private partnerships.

Future Developments

NEC will collaborate with the Defense Innovation Science and Technology Institute, aiming to develop this foundational model, demonstrate its effectiveness, and complete its construction by fiscal year 2027.

Moving forward, NEC will continue to promote research and development aimed at high-precision visualization and a better understanding of the underwater environment. Furthermore, by contributing to both the strengthening of defense capabilities and the development of the maritime industry through the implementation of this foundational model—and by accelerating dual-use applications—NEC will contribute to the realization of a safe, secure, and sustainable society.

Notes:

(*1) The Innovative Breakthrough Research program is conducted under managers recruited from outside Japan’s Ministry of Defense. This is part of an initiative by the Agency for Defense Equipment to create innovative and groundbreaking functions and technologies that fundamentally strengthen defense capabilities or transform society.

(*2)

“cotomi” is a core AI technology developed by NEC.

https://group.nec/jp/ja/solutions/ai/llm/ (Japanese website)

About NEC

The NEC Group leverages technology to create social value and promote a more sustainable world where everyone has the chance to reach their full potential. NEC Corporation was established in 1899. Today, the NEC Group’s approximately 110,000 employees utilize world-leading AI, security, and communications technologies to solve the most pressing needs of customers and society.

NVIDIA Unveils Nemotron 3.5 Lightning: 30 Billion Open‑Source AI Model for Faster Agentic Workflows

NVIDIA has just launched Nemotron 3.5 Lightning, a 30-billion-parameter open-source AI model optimized for agentic workflows. It delivers up to 4× faster output and 30% quicker task completion compared to peers, while being lightweight enough to run locally on a single GPU — even on laptops and desktops.

Key Highlights

• Architecture: Mixture-of-Experts (MoE) hybrid with Mamba-2 + Attention layers.
• Parameters: 30B total, 3B active per token (efficient activation).
• Performance: Up to 4× faster output, 30% faster agentic task completion, PinchBench accuracy: 86%.
• Deployment: Runs on RTX PCs, Jetson devices, DGX Spark, workstations, data centers, and cloud.
• Open-source availability: Free to download via HuggingFace, NVIDIA NGC, and build.nvidia.com.
• Use cases: Code review, cybersecurity monitoring, billing queries, tool calls, and multi-agent workflows.

Comparison with Other Models

Model	Parameters	Specialization	Performance	Deployment
Nemotron 3.5 Lightning	30B (3B active)	High-volume agent execution	4× faster, 30% quicker tasks	Single GPU, edge, cloud
Nemotron 3 Ultra	Larger frontier model	Planning & orchestration	Advanced reasoning	Data center/cloud
Qwen3.6 35B	35B	General-purpose	Slower task completion	Cloud-heavy

Technical Features

• Multi-token prediction: MTP + speculative decoding for concurrency.
• Draft models: DSpark (DGX Spark inference), DFlash (workload evaluation).
• Formats: BF16 (full precision) and NVFP4 (optimized inference).
• Context length: Up to 1M tokens, enabling long workflows.
• Supported languages: English, Spanish, French, German, Italian, Japanese, plus coding languages.

Strategic Context

• Open-weight AI: Jensen Huang emphasized open-weight AI as critical for innovation, lowering entry barriers for enterprises while driving GPU demand.
• Commercial use: Free to use commercially, unlike proprietary subscription-based AI services.
• Concerns: National security risks and IP leakage remain, but NVIDIA positions Nemotron as a cornerstone for agentic AI ecosystems.

IndiaAI backed Avataar Launches Varya, India’s Distilled Video AI Model, Delivering Frontier‑quality Video Generation at 10x Lower Cost

India AI mission backed AI-native transformation company, Avataar, has launched Varya, India’s first distilled video AI model under the IndiaAI Mission, promising frontier‑quality video generation at up to 10x lower cost.

Launch Overview

• Event: Press launch in New Delhi, June 12, 2026
• Key presence: Shri S. Krishnan, Secretary, MeitY, alongside Avataar leadership
• Backed by: IndiaAI Mission and Peak XV Partners
• Objective: Deliver efficient, culturally aware video AI for India’s billion‑plus users

What Makes Varya Unique

• Distilled architecture: Compresses video generation from 50 steps to 4, maintaining comparable quality
• Efficiency: Generates video at ₹0.48 per second (~$0.005), 20x cheaper than global rivals
• Scale: 14‑billion parameter model optimized for India’s contexts
• Speed: On NVIDIA H200 GPUs, a 5‑second 720p clip takes 45 seconds vs. 1,230 seconds for Alibaba’s Wan 2.2
• Accessibility: Available via IndiaAI Kosh for developers to self‑host or adapt

Cultural Context & Applications

• Education: Teachers creating visual lessons in rural classrooms
• MSMEs: Affordable product ads and digital storytelling
• Citizen services: Public information through video
• Commerce: E‑commerce and marketing campaigns

Leadership Statements

• S. Krishnan, MeitY: “The launch of one of the foundational models supported under the IndiaAI Mission marks a significant milestone in India’s AI journey. Varya reflects our commitment to building indigenous AI capabilities and fostering a vibrant deep‑tech ecosystem.”
• Sravanth Aluru, CEO, Avataar: “India’s AI opportunity will not be defined only by the largest models. It will also be defined by the most efficient models. For a country of 1.4 billion people, affordability is not a feature, it is a prerequisite.”

Global Comparison

Model	Generation Steps	Cost per Second	Cultural Context
Varya	4	₹0.48 ($0.005)	India‑specific, culturally rich
Wan 2.2	50	Higher (benchmark)	Generic global
Veo / Runway	50+	$0.10+	Global, less India‑specific

Risks & Opportunities

• Opportunities: Democratizes video AI for education, MSMEs, and governance; positions India as a global leader in frugal AI innovation
• Risks: Global competition may push rapid iterations; infrastructure dependency on subsidized compute; cultural bias risks if training data isn’t updated

In essence, Varya is not just a technical breakthrough but a strategic milestone — redefining India’s AI ambition by proving that efficiency, affordability, and cultural relevance can rival global frontier models.

NPCI Launches FiMI: India’s First Finance-Specific AI Language Model for Payment

National Payments Corporation of India (NPCI) today released FiMI (Finance Model for India), a domain-specific language model built for India’s payments ecosystem, at the India–AI Impact Summit 2026. Developed in-house by NPCI, FiMI currently powers NPCI’s UPI Help Assistant and is designed to address limitations observed in general-purpose large language models when applied to high-scale, high-trust payment workflows.

FiMI has been built to natively understand the complexity of Indian payment systems, including UPI, covering transaction dispute handling, mandate lifecycle management, and regulatory and ecosystem queries. The model went through continuous pre-training and was further fine-tuned on Indian financial and synthetically generated payments data, enabling accurate reasoning, structured tool invocation, and robust multilingual support. This payments-native design allows FiMI to operate reliably in real-world, high-volume transaction environments where accuracy, consistency, and trust are critical.

FiMI is already deployed at a national scale through NPCI’s UPI Help Assistant launched as pilot initiative, an AI-powered conversational support system for UPI users. This assistant leverages an agentic AI framework that enables multi-step reasoning to respond to payment related queries, support grievance redressal, and assist with mandate management. The assistant currently supports English, Hindi, Telugu, and Bengali, with additional Indian languages planned to be introduced within the next 6 to 8 months to further expand accessibility.

Having successfully developed and validated FiMI in a production environment, NPCI has released a detailed technical paper on the arXiv platform outlining the data curation process, training methodology, and evaluation results. This publication reflects NPCI’s role as a public digital infrastructure institution and its commitment to advancing transparent, India-centric, and trustworthy AI for the payments ecosystem.

Through this technical publication, NPCI is sharing key learnings from building and operating AI systems at a national scale, contributing to broader research and enabling informed development of domain-specific AI systems for financial infrastructure.

Commenting on the announcement, Vishal Kanvaty, Chief Technology Officer, NPCI said, “FiMI reflects NPCI’s commitment in developing financial language models purpose-built for India’s payments ecosystem. Having successfully built and validated the model, we are now sharing our technical learnings through a detailed paper to contribute to responsible innovation, while strengthening trust, transparency, and collaboration across the ecosystem.”

Building on FiMI, NPCI is exploring next-generation model architectures, including Mixture-of-Experts, alongside the expansion of multilingual capabilities to cover a wider range of Indian languages. These efforts are currently in the research and experimentation phase and are focused on improving scalability, efficiency, and domain intelligence, while preserving the governance, reliability, and trust requirements essential for national-scale payment systems.

The detailed technical paper can be accessed here.

AI Designs Viruses That Kill Bacteria—A New Frontier in Synthetic Biology

In a stunning leap for synthetic biology, scientists have used artificial intelligence to design viruses that can infect and kill bacteria—ushering in a new era of programmable life forms and potentially revolutionizing medicine.

Researchers at Stanford University and the Arc Institute trained an AI model named Evo on over 2 million bacteriophage genomes. The goal? To teach the system how nature builds viruses that target bacteria. Evo didn’t just remix existing genetic material—it generated 302 entirely new viral genomes, many of which had never existed in nature.

Of those, 16 assembled into fully functional viruses that successfully infected and destroyed E. coli bacteria in lab tests. This marks the first time AI has been used to design complete, working viruses from scratch.

“We’re not just accelerating evolution—we’re directing it,” said one of the lead researchers. “This opens the door to custom-built phages that could target antibiotic-resistant bacteria with surgical precision.”

Why This Matters

• Antibiotic resistance is one of the biggest threats to global health, with superbugs killing over a million people annually.
• Phage therapy, which uses viruses to kill bacteria, has long been seen as a promising alternative—but finding the right phage is slow and unpredictable.
• AI could dramatically speed up the discovery and design of targeted phages, potentially enabling personalized treatments for infections.

The Ethical Frontier

While the study focused solely on bacteriophages and excluded viruses that infect humans, the implications are profound. Experts warn that AI-designed viruses could behave unpredictably in complex ecosystems. There are also concerns about biosecurity and the potential misuse of such technology.

“We need robust oversight and ethical frameworks,” said a bioethicist not involved in the study. “This is powerful tech, and with great power comes great responsibility.”

What’s Next?

• The team plans to expand Evo’s capabilities to design phages for other bacterial strains, including those responsible for hospital-acquired infections.
• There’s growing interest in using AI to design viruses for agriculture, microbiome engineering, and environmental cleanup.

This breakthrough isn’t just about killing bacteria—it’s about reimagining what life can be. With AI as a co-creator, biology may no longer be bound by the slow march of evolution. It’s entering the age of intelligent design.

Top 10 Use Cases of Model Context Protocol (MCP)

Artificial intelligence has made huge strides in the last few years, but one challenge has always remained. Most AI models forget everything the moment a session ends. You can ask it something today, come back tomorrow, and it acts like you never spoke before. This is where Model Context Protocol (MCP) changes the game.

MCP allows AI models to retain and refer back to past interactions, settings, and preferences. In simpler terms, it helps AI "remember" what happened before, so responses can be more accurate and meaningful over time. This feature is now being integrated across a wide range of industries.

Let’s look at ten practical and impactful ways MCP is being used today. The content below, authored by Dr Ananth G of DaveAI, explores practical applications of MCP across industries such as healthcare, education, gaming, finance, and more, highlighting how this breakthrough is shaping the future of AI.

Dr. Ananth 

1. Conversational AI That Actually Remembers

Chatbots and virtual assistants are often helpful, but their biggest weakness is forgetfulness. Without context, they cannot follow up on past conversations or respond in a more human-like way. MCP enables AI to maintain memory across sessions, creating a smoother and more personalized interaction.

For example, if a customer reports a billing issue and comes back a few days later, the AI can pick up where it left off. This removes the frustration of repeating details and makes the conversation feel more natural and efficient.

2. Smarter Personal Assistants

Virtual assistants like Siri and Alexa are useful, but they can still feel limited. MCP upgrades their intelligence by allowing them to learn your habits, schedule, and preferences over time.

If you usually play music at 7 AM or check the weather before leaving for work, MCP lets your assistant pick up on those patterns. Over time, it can become more proactive, offering reminders and actions without needing constant instructions.

3. Adaptive Learning Platforms

Online learning platforms often lack memory across sessions. A student might get help with a topic today, only for the system to start fresh the next day. MCP changes that by allowing the platform to remember learning progress, strengths, weaknesses, and even preferred learning styles.

This leads to more personalized and effective education. Tutors, whether human or AI, can provide targeted support based on a learner’s journey, not just their most recent activity.

4. Enhanced Legal and Financial Advisories

In legal and financial services, having context is essential. An AI assistant helping with contracts or investment planning becomes far more useful when it remembers past discussions, documents reviewed, and user preferences. MCP allows AI to hold onto these important details. That way, it can offer consistent advice, track legal cases over time, and maintain continuity in financial planning. The result is a more competent and reliable digital advisor.

5. More Nuanced Content Creation

Writers who use AI tools often find themselves re-entering the same instructions each time. Whether it is tone, structure, or content goals, MCP helps AI remember those creative choices across sessions. If you are working on a blog series, an AI tool using MCP can recall the style and content from earlier posts. This ensures consistency and saves time, especially for marketing teams, journalists, and solo creators working on long-term projects.

6. Consistent Character Development in Games

Games are becoming more immersive, and players expect more from non-playable characters (NPCs). Instead of scripted reactions, MCP lets NPCs respond based on a player’s past actions. If a player forms an alliance or betrays a character early on, that NPC can react differently later in the game. This creates deeper, more realistic storytelling. MCP turns video game interactions into ongoing relationships, making the experience far more engaging.

7. Enterprise Workflows with Memory

Many businesses now use AI to help with project management, document handling, and internal communication. MCP brings continuity into these workflows by allowing AI systems to remember the scope of a project or the decisions made in earlier meetings. This reduces mistakes, avoids repetition, and makes it easier to stay aligned with long-term goals. Project managers and teams benefit from AI that understands the full context of their work over time, not just what happened today.

8. Medical Assistants and Patient History

Healthcare depends heavily on history. A patient’s treatment, diagnosis, and test results must all be tracked carefully. MCP allows AI systems to retain this context, making them more reliable helpers in medical settings. When used in virtual assistants or diagnostic tools, MCP helps track symptoms, past appointments, and treatment responses. Doctors can receive AI support that is better informed and more accurate, ultimately leading to better care.

9. Dynamic Travel and Itinerary Planning

Planning a trip involves dozens of details. AI travel assistants can use MCP to remember preferences like hotel type, preferred airlines, meal choices, or even sleep schedules. This helps the system adapt when changes are needed. If you miss a flight or want to adjust your schedule, the AI can replan everything with your previous choices in mind. It is like having a travel agent who knows you well, without having to explain yourself each time.

10. Long-Term Coaching and Therapy Support

Coaching apps and mental health platforms are increasingly turning to AI for support. MCP enables these tools to offer a continuous experience by remembering emotional tone, session goals, and prior conversations. This is especially useful in therapy, where a consistent understanding of the user is key. MCP ensures that each session builds on the last, leading to stronger trust and more meaningful guidance. Whether it is career coaching or emotional support, AI becomes more helpful when it truly remembers.

Final Thoughts

Model Context Protocol is not just a technical feature. It represents a significant shift in how AI systems interact with humans. By remembering what came before, AI becomes more useful, more intuitive, and far less frustrating to use. From helping businesses stay on track to improving medical advice and creating better games, MCP is quietly transforming how we use technology in everyday life. As this capability becomes more widespread, we can expect AI to feel less like a tool and more like a thoughtful assistant, one that listens, remembers, and truly understands.

Next-Level AI Training: Quantum Computing Fine-Tunes Billion-Parameter Model

Chinese researchers have achieved a global first by using a real quantum computer to fine-tune an Al model with one billion parameters. The experiment was conducted on Origin Wukong, China's third-generation superconducting quantum computer with 72 qubits.

Workers calibrate and install the China's independently developed third-generation superconducting quantum computer. Photo:Courtesy: Anhui Quantum Computing Engineering Research Center

This breakthrough led to an 8.4% improvement in training performance while reducing the number of parameters by 76%. The Al model also showed better results in specific tasks-when trained on mental health conversation data, it made 15% fewer mistakes, and in a math problem-solving test, its accuracy jumped from 68% to 82%.

The fine-tuning process traditionally requires high computing power, but quantum computing offers unique Chinese researchers have achieved a global first by using a real quantum computer to fine-tune an Al model with one billion parameters. The experiment was conducted on Origin Wukong, China's third-generation superconducting quantum computer with 72 qubits.

This breakthrough led to an 8.4% improvement in training performance while reducing the number of parameters by 76%. The Al model also showed better results in specific tasks-when trained on mental health conversation data, it made 15% fewer mistakes, and in a math problem-solving test, its accuracy jumped from 68% to 82%.

The fine-tuning process traditionally requires high computing power, but quantum computing offers unique advantages. By leveraging superposition and entanglement, quantum computers can explore vast combinations of parameters simultaneously, making Al training faster and more efficient.

This development could be a game-changer for Al training, reducing computational costs and improving model efficiency.

Experts have reacted with cautious optimism to China's breakthrough in using a quantum computer to fine-tune a billion-parameter AI model.

Some experts remain skeptical, pointing out that while the results are promising, the research is still in the demonstration phase and lacks peer-reviewed validation.

Amazon Introduces New Gen AI Model for Building Voice Apps & Agents

Today, Amazon.com Inc (NASDAQ: AMZN) introduced Amazon Nova Sonic, a new foundation model that unifies speech understanding and speech generation into a single model, to enable more human-like voice conversations in artificial intelligence (AI) applications. Available in Amazon Bedrock via a new bi-directional streaming API, the model simplifies the development of voice applications, such as customer service call automation and AI agents across a broad range of industries, including travel, education, healthcare, entertainment, and more.

“From the invention of the world’s best personal AI assistant with Alexa, to developing AWS services like Connect, Lex, and Polly that are used across a wide range of industries, Amazon has long believed that voice-powered applications can make all of our customers’ lives better and easier,” said Rohit Prasad, SVP of Amazon Artificial General Intelligence. “With Amazon Nova Sonic, we are releasing a new foundation model in Amazon Bedrock that makes it simpler for developers to build voice-powered applications that can complete tasks for customers with higher accuracy, while being more natural, and engaging.”

Traditional approaches to building voice-enabled applications involve complex orchestration of multiple models, such as speech recognition to convert speech to text, large language models (LLMs) to understand and generate responses, and text-to-speech to convert text back to audio. This fragmented approach not only increases development complexity but also fails to preserve crucial acoustic context and nuances like tone, prosody, and speaking style that are essential for natural conversations.

Nova Sonic solves these challenges through a unified model architecture that delivers speech understanding and generation, without requiring a separate model for each of these steps. This unification enables the model to adapt the generated voice response to the acoustic context (e.g. tone, style) and the spoken input, resulting in more natural dialog. Nova Sonic even understands the nuances of human conversation, including the speaker’s natural pauses and hesitations, waiting to speak until the appropriate time, and gracefully handling barge-ins. It also generates a text transcript for the user’s speech, enabling developers to use that text to call specific tools and APIs for building voice-enabled AI agents (e.g., an AI-powered travel agent that can book flights by retrieving up to date flight information). These capabilities, along with its lightning-fast inference, make voice applications powered by Nova Sonic more natural and useful.

State-of-the-art accuracy and quality

Nova Sonic has been rigorously tested against a wide range of industry standard benchmarks for speech understanding and generation, demonstrating exceptional quality and accuracy for human-like, real-time voice conversations.

The model excels in natural dialog handling, seamlessly understanding and adapting to pauses, hesitations, and interruptions while maintaining conversational context throughout the interaction. This capability contributed to strong performance for overall quality and accuracy in turn-taking tests.

Nova Sonic demonstrates strong performance on overall conversation quality compared to other models in the industry, which at this time include a select few with similar real-time conversational speech capabilities, such as OpenAI's GPT-4o (Realtime) and Google Gemini Flash 2.0 (available via Gemini’s experimental live API). For example, single-turn dialogs in its American English masculine-sounding voice achieved a 51.0% and 69.7% win-rate against OpenAI’s GPT-4o (Realtime) and Google’s Gemini Flash 2.0 respectively, based on the Common Eval data set. Likewise, Nova Sonic’s American English feminine-sounding voice scored 50.9% and 66.3% win-rate against OpenAI’s GPT-4o (Realtime) and Google’s Gemini Flash 2.0 respectively on the same data set. Nova Sonic also exceeds performance for its British English feminine-sounding voice, scoring a 58.3% win-rate against OpenAI’s GPT-4o (Realtime).

Since recognizing spoken words is critical in generating accurate responses, measuring Nova Sonic's speech recognition accuracy in terms of word error rate (WER) across a wide range of languages, dialects, and accents is also critical. On the Multilingual LibriSpeech, Nova Sonic achieved a WER of 4.2%, which is 36.4% relative lower than OpenAI's GPT-4o Transcribe model, when averaged across English, French, Italian, German and Spanish.

On English utterances of the Multilingual LibriSpeech (MLS) data set, it has 24.2% relative lower WER compared to OpenAI’s GPT-4o Transcribe model.

Nova Sonic is also robust to noisy conditions, with 46.7% relative lower WER for English compared to OpenAI’s GPT-4o Transcribe model measured on Augmented Multi Party Interaction (AMI) meeting benchmark that consists of real-world noisy and multi-speaker interactions.

Tool-use for function calling and agentic workflows

Nova Sonic also supports tool-use for applications—like customer service call automation—that require the responses to be factually grounded in enterprise data, such as pricing plans, available inventory, and schedule availability. Nova Sonic’s native tool-use also enables the model to resolve complex customer queries and complete tasks on behalf of customers, for example, “make a reservation” or “find alternate flights.”

Multiple native voices and speaking styles

Nova Sonic supports three expressive voices, including both masculine-sounding and feminine-sounding voices now generally available in English, and supports speech generation in different English accents including American and British. Support for additional languages and accents will be coming soon.

Industry-leading speed and price performance

Nova Sonic delivers an average customer-perceived latency of 1.09 seconds from the time the customer is done talking to the time the system generates the first speech response. This is compared to 1.18 seconds for OpenAI’s GPT-4o (Realtime), and 1.41 seconds for Google’s Gemini Flash 2.0 (available via Gemini’s experimental live API), per benchmarking by Artificial Analysis.

Nova Sonic is the most cost-efficient model in the industry, when compared to models that have similar functionality of real-time speech conversations and have public pricing available. For example, it is nearly 80% less expensive than OpenAI’s GPT-4o (Realtime).

Amazon Nova Sonic is helping companies drive better customer satisfaction and productivity

ASAPP empowers enterprise customers’ contact centers to deliver unmatched customer service through GenerativeAgent, a fully conversational generative Al voice agent. “At ASAPP, we are focused on using generative AI to deliver reliable, secure, and high-performing solutions for improving customer service in contact centers. We’ve been particularly impressed by Amazon Nova Sonic’s highly accurate speech understanding capabilities which allow for more natural voice interactions and precise dialog handling over telephony,” said Nirmal Mukhi, VP of AI Engineering at ASAPP. “We’re excited to continue using Nova Sonic to deliver secure, high-quality, and precise conversations that meet the demands of enterprise contact centers.”

Education First (EF) is a leader in international education through its networks of schools and offices in over 50 countries. “Amazon Nova Sonic enables EF students to practice new vocabulary and refine their pronunciation in a dynamic learning environment, while the interactive nature of the model allows students to receive immediate feedback on their pronunciation attempts, contributing to a more efficient and effective learning process. The model is capable of accurately understanding non-native English speakers with a variety of accents. We were also impressed with the barge-in feature of Nova Sonic, where the model quickly reacts to interruptions,” said Tim Hesse, VP of AI and Data at EF. “The scalability and reliability of the technology will allow us to expand our capacity to serve a larger student population simultaneously, without compromising the quality of instruction.”

Stats Perform is a sports data and AI technology provider, serving global media organizations, betting operators, and professional sports teams. “At Stats Perform, our goal is to empower the world’s top sports broadcasters, media, federations and teams with magic in the detail of our vast live and historical Opta sports dataset, to help them win audiences, customers and trophies. With the Opta AI Chat they can generate unique, accurate, and contextual responses, driven by live data insights with remarkable speed, in multiple formats and languages, to find a winning analytical or storytelling edge,” said Mike Perez, Chief Operating Officer at Stats Perform. “We’ve been testing Amazon Nova Sonic and have been particularly impressed by the system's low latency, which enables near-instantaneous responses even to complex queries of our model, creating a seamless user experience that turns human experts into superhuman experts. The intuitive prompting capability and ease of setup have exceeded our expectations, making implementation simple. Overall, Nova Sonic has proven to be a fantastic solution.”

Amazon is committed to the responsible development of artificial intelligence

Amazon Nova models are built with integrated safety measures and protections. The company has launched AWS AI Service Cards for Nova models, offering transparent information on use cases, limitations, and responsible AI practices.

To get started with Amazon Nova models, visit: https://aws.amazon.com/nova/

To learn more, visit: About Amazon for details on today’s announcement.

Foxconn Launches Its First AI Large Model 'FoxBrain'

Hon Hai Research Institute announced today the launch of the first Traditional Chinese Large Language Model (LLM), setting another milestone in the development of Taiwan’s AI technology with a more efficient and lower-cost model training method completed in just four weeks.

The institute, which is backed by Hon Hai Technology Group (“Foxconn”) (TWSE:2317), the world’s largest electronics manufacturer and leading technological solutions provider, said the LLM – code named FoxBrain – will be open sourced and shared publicly in the future. It was originally designed for applications used in the Group’s internal systems, covering functions such as data analysis, decision support, document collaboration, mathematics, reasoning and problem solving, and code generation.

FoxBrain not only demonstrates powerful comprehension and reasoning capabilities but is also optimized for Taiwanese users' language style, showing excellent performance in mathematical and logical reasoning tests.

"In recent months, the deepening of reasoning capabilities and the efficient use of GPUs have gradually become the mainstream development in the field of AI. Our FoxBrain model adopted a very efficient training strategy, focusing on optimizing the training process rather than blindly accumulating computing power,” said Dr. Yung-Hui Li, Director of the Artificial Intelligence Research Center at Hon Hai Research Institute. ”Through carefully designed training methods and resource optimization, we have successfully built a local AI model with powerful reasoning capabilities."

The FoxBrain training process was powered by 120 NVIDIA H100 GPUs, scaled with NVIDIA Quantum-2 InfiniBand networking, and finished in just about four weeks. Compared with inference models recently launched in the market, the more efficient and lower-cost model training method sets a new milestone for the development of Taiwan's AI technology.

FoxBrain is based on the Meta Llama 3.1 architecture with 70B parameters. In most categories among TMMLU+ test dataset, it outperforms Llama-3-Taiwan-70B of the same scale, particularly exceling in mathematics and logical reasoning (For TMMLU+ benchmark of FoxBrain, please refer to Fig.1). The following are the technical specifications and training strategies for FoxBrain:

• Established data augmentation methods and quality assessment for 24 topic categories through proprietary technology, generating 98B tokens of high-quality pre-training data for Traditional Chinese
• Context window length: 128 K tokens
• Utilized 120 NVIDIA H100 GPUs for training, with total computational cost of 2,688 GPU days
• Employed multi-node parallel training architecture to ensure high performance and stability

Used a unique Adaptive Reasoning Reflection technique to train the model in autonomous reasoning test results, FoxBrain showed comprehensive improvements in mathematics compared to the base Meta Llama 3.1 model. It achieved significant progress in mathematical tests compared to Taiwan Llama, currently the best Traditional Chinese large model, and surpassed Meta's current models of the same class in mathematical reasoning ability. While there is still a slight gap with DeepSeek's distillation model, its performance is already very close to world-leading standards.

FoxBrain's development – from data collection, cleaning and augmentation, to Continual Pre-Training, Supervised Finetuning, RLAIF, and Adaptive Reasoning Reflection – was accomplished step by step through independent research, ultimately achieving benefits approaching world-class AI models despite limited computational resources. This large language model research demonstrates that Taiwan's technology talent can compete with international counterparts in the AI model field.

Although FoxBrain was originally designed for internal group applications, in the future, the Group will continue to collaborate with technology partners to expand FoxBrain's applications, share its open-source information, and promote AI in manufacturing, supply chain management, and intelligent decision-making.

During model training, NVIDIA provided support through the Taipei-1 Supercomputer and technical consultation, enabling Hon Hai Research Institute to successfully complete the model pre-training with NVIDIA NeMo. FoxBrain will also become an important engine to drive the upgrade of Foxconn’s three major platforms: Smart Manufacturing. Smart EV. Smart City.

The results of FoxBrain is scheduled to be shared for the first time at a major conference during NVIDIA GTC 2025 Session Talk “From Open Source to Frontier AI: Build, Customize, and Extend Foundation Models” on March 20.

Alibaba Open-Sources Its AI Video Model Wan2.1 Series

Alibaba is making waves in the AI world by launching an open-source version of its video and image-generating AI model, Wan 2.1. This move is set to intensify competition in China's AI market, especially following DeepSeek's recent launch of its own open-source models.

Alibaba Cloud announced the open source release of four models in its Wan2.1 series of large video generation models. As an open source, it will be open to global academia, researchers, and commercial organizations for use, further promoting innovation and inclusiveness of artificial intelligence (AI) technology.

Alibaba's AI models, particularly the Qwen 2.5-Max and Wan 2.1, are making significant strides in the AI landscape. Alibaba Cloud is one of the first global technology companies to open source its own large-scale AI models, and as early as August 2023, it launched its first open source model Qwen (Qwen-7B). 

The open-source community is unstoppable! 🌟 To celebrate #OpenSourceWeek, we’re excited to open-source Wan2.1—the #1 Hugging Face foundation video model and the only open-source video generation model in the VBench top 5. Let’s keep on supporting open-source together!🤝

Check… https://t.co/PrkFv2iWDT

— Alibaba Group (@AlibabaGroup) February 26, 2025

Wan 2.1 is designed to generate highly realistic visuals and has already secured a top ranking on VBench, a leaderboard for video generative models. Alibaba has released Wan 2.1, each capable of generating images and videos from text and image input. These models are available globally on Alibaba Cloud's ModelScope and HuggingFace platforms.

In addition to Wan 2.1, Alibaba has also introduced a preview version of its reasoning model, QwQ-Max, which it plans to make open source upon the full release. This strategic move aligns with Alibaba's broader AI ambitions, as the company has announced plans to invest at least $52 billion over the next three years to bolster its cloud computing and AI infrastructure.

The Qwen 2.5-Max model is part of Alibaba's open-source Qwen series and is designed to process long, complex queries and engage in nuanced conversations. It has been benchmarked against models like OpenAI's GPT-4, DeepSeek-V3, and Meta's Llama-3.1-405B, and has shown superior performance in several areas.

This open-source initiative is expected to foster innovation, lower barriers to entry, and position Alibaba as a formidable player in the AI space.

Alibaba in its announcement press release said — Training video-based models requires huge computing resources and a large amount of high-quality training data. Open source helps lower the barrier to entry for more companies to use AI, enabling them to create high-quality visualization content that meets their needs in a cost-effective manner.

Text prompt: A man is performing professional diving moves on a diving platform. In the panoramic shot, he is wearing red swimming trunks and his body is upside down, with his arms extended and his legs together. The camera moves down and he jumps into the water, splashing. In the background is a blue swimming pool.)

Among them, the T2V-14B model is more suitable for generating high-quality visual effects with rich motion dynamics, while the T2V-1.3B model strikes a balance between generation quality and computing power, making it an ideal choice for developers for secondary development and academic research. For example, the T2V-1.3B model allows users to generate a 5-second, 480p resolution video in about 4 minutes using only an ordinary laptop.

"India is Not a Poor Country...We Must Have the Courage to Build Our Own Foundational AI Models”: Google Deepmind’s Dr. Manish Gupta at ABP Network’s Ideas of India 2025

• “AI can enhance human abilities, but humans will always be capable of far more. The future is not about AI replacing us—it is about AI making us better,” Dr. Manish Gupta.
• “Our mission is to build AI responsibly for humanity, ensuring it benefits everyone—not just the privileged few,” Dr. Manish Gupta.

“India is not a poor country anymore, and we must take bold steps in developing our foundational AI models.” Dr. Manish Gupta, Senior Director at Google DeepMind, who leads AI research teams across India and Japan, shared his perspective at ABP Network’s Ideas of India 2025 on the crucial role of AI in transforming India and its power in solving global challenges.

Speaking on the topic ‘Transforming India with AI: Why We Need More Data’, Dr. Manish Gupta, said,

Some assume that building foundational AI models requires hundreds of millions of dollars, but we must have the courage to take on challenges that have never been attempted before. AI is an old discipline, almost as old as computing itself. The reason ChatGPT and similar models took off is the innovation of transformer architecture, which has been key to foundational AI models.

Adding to the conversation, Dr. Manish Gupta, said,

AI is accelerating progress on a range of critical problems, from drug discovery and plastic pollution to structural biology. Innovations like AlphaFold are helping us solve complex scientific challenges faster than ever before.

Sharing his views on AI’s role in language and culture, Dr. Gupta added, “We are focused on building the best multimodal representation for 100+ Indic languages, ensuring AI development is inclusive and supports India's linguistic diversity. Multicultural AI models will shape the future of technology. Our mission is to build AI responsibly for humanity. The world is excited about AI, but we must ensure it benefits everyone, not just the privileged few. AI expertise should be widespread and accessible to all."

Talking about AI’s future capabilities, ethics and emotional understanding, he commented, “AI is biased because it is built by humans, and we carry our biases into these systems. The reality is that AI still struggles to understand emotions like love. Companies like Google are working on AI principles to mitigate bias and ensure responsible AI development. AI can become a better version of human beings, like AI can help artists explore new genres, but when it comes to competition, humans will always win. AI can hallucinate, but human creativity and intuition will always have the upper hand.”

The ABP Network’s Ideas of India 2025, centred on the theme ‘Humanity’s Next Frontier’, will convene thought leaders and innovators to explore the challenges and opportunities in India’s ascendance in a rapidly changing world. In the face of climate change, geopolitical conflicts, and technological advancements like AI, the summit delved into India’s role as both an ancient civilisation and a demographic powerhouse in shaping the future. The two-day summit brings together a confluence of ideas from global thought leaders, intellectuals, and change-makers, covering transformative possibilities in science, medicine, social contracts, and global leadership, with experts from diverse fields offering bold visions of a better, more sustainable world for all.

About ABP Network

An innovative media and content creation company, ABP Network is a credible voice in the broadcast and digital spheres, with a multi-language portfolio of news channels reaching 535 million individuals in India across YouTube, websites, apps, and social media platforms. ABP Studios, which comes under the purview of ABP Creations – the content innovation arm of the network – creates, produces, and licences original, groundbreaking content outside of news. ABP Network is a group entity of ABP, established almost 100 years ago, and continues to dominate as a leading media company.

Chinese Startup's Cheaper AI Model Creates Ripples Causing US Stock Market to Drop by $1 Trillion in Single Day

China's DeepSeek AI has recently made headlines by shaking up the U.S. stock market. Here's a brief overview of what happened:

DeepSeek AI, a Chinese artificial intelligence startup, announced a breakthrough in AI technology with its R1 model, which can rival the performance of leading U.S. AI models like OpenAI's GPT-4 but at a fraction of the cost. This announcement caused a significant drop in the stock market, particularly affecting tech giants like Nvidia, which saw a historic $589 billion loss in market value in a single day. The overall impact on the tech-heavy Nasdaq index was a drop of nearly $1 trillion in global market value.

On 10 January 2025, DeepSeek released its first free chatbot app, based on the DeepSeek-R1 model, for iOS and Android. By 27 January, DeepSeek-R1 had surpassed ChatGPT as the most-downloaded free app on the iOS App Store in the United States, causing Nvidia's share price to drop by 18%. Besides United States, the Deepseek effect could also be seen in Japan as SoftBank stock price also fell by 8% during trading hours in Tokyo.

The key factors contributing to these market reaction include:

• Cost Efficiency: DeepSeek's AI model was developed for $5.6 million, much less than the billions spent by U.S. companies on similar models.
• Technological Leap: DeepSeek's model challenges the dominance of U.S. tech firms and raises questions about the sustainability of high capital expenditures in AI development.
• Market Panic: The sudden rise of a competitive AI model from China led to investor uncertainty and a sell-off in tech stocks.

This event has sparked debates about the economic and geopolitical competition between the U.S. and China in the AI industry. It's indeed a significant development in the tech world!

Headquartered in Hangzhou, Zhejiang, DeepSeek is a Chinese AI startup founded by Liang Wenfeng in May 2023. As for investors, DeepSeek is funded by the Chinese hedge fund High-Flyer, which was co-founded by Liang Wenfeng. High-Flyer focuses on developing and using AI trading algorithms.

The 20-months old Chinese start-up has quickly gained attention with its DeepSeek-V3 and DeepSeek-R1 models, which have been praised for their efficiency and cost-effectiveness. Here are some key points about DeepSeek:

Cost-Effective Development: DeepSeek-V3 was trained using approximately 2,000 Nvidia H800 chips over 55 days, costing around $5.58 million. This is significantly less than the billions spent by other companies on similar models.

Advanced Capabilities: DeepSeek-R1 focuses on logical inference, mathematical reasoning, and real-time problem-solving, achieving performance comparable to OpenAI's o1 model.

Open Source: DeepSeek's models are available for free, making them accessible to a wider audience.

Global Impact: The launch of DeepSeek has caused a significant reaction in global markets, particularly affecting U.S. tech stocks.

DeepSeek's innovative approach to AI development has sparked discussions about the future of AI technology and its potential to democratize access to advanced AI tools.