XTHE

Samsung Electronics has started early trials of EUV lithography at its Taylor, Texas, foundry. Equipment testing begins soon.  

Scheduled for March 2026, tests are being prepared for 2nm chip production. Samsung brings GAA manufacturing to the US, competing with TSMC.  

Important Information About The Trial And Production Plan Includes: 

• Trial timeline: EUV machine trials will start in early 2026.  

• Production focus: Taylor will shift from older processes to 2nm technology to support high-performance AI chips for clients such as Tesla.  

• Early tech adoption: EUV pellicles at Texas aim to boost yield and efficiency.  

• Initial reports expected 2nm production by late 2026, but full-scale mass production may shift to early 2027 due to process setbacks.  

• Strategic change: Samsung will use Taylor, which is larger than the combined Hwaseong and Pyeongtaek Korean sites, to fully serve AI chip makers.  

The Texas facility has a temporary Certificate of Occupancy, letting Samsung install and test equipment. These trials help stabilize 2nm yields to meet strict requirements.  

Samsung is preparing for a major milestone at its US semiconductor factory in March 2026. The company will test extreme ultraviolet lithography equipment at its Taylor, Texas, plant. This move brings Samsung closer to producing advanced chips, including Tesla’s next-generation chips.  

Preparations are underway for advanced chip production at Samsung’s Taylor plant.  

Last month, reports said Samsung would install its first manufacturing equipment and launch trial operations at Taylor in March 2026. The company plans staged equipment installation and full operations in the second half of 2026.  

The report indicates Samsung may seek temporary occupancy authorization from authorities for Plant 1, enabling use before construction ends if requirements are met. Engineers from headquarters are at Taylor to rapidly stabilize production yields.  

Construction at the Taylor plant involves about 7,000 workers daily. Approximately 1,000 are building a 6-storey office, expected to finish in the second half of 2026. The facility covers about 4.8 million square meters, larger than the semiconductor complexes in Pyeongtaek and Hwaseong. The plant will focus on advanced processes, including 2nm technology. ASML supplies the essential EUV equipment.  

Samsung has secured initial orders from Taylor, producing the autonomous-driving chips AI-5 and AI-6 for Tesla. If standards are met, Samsung may receive more Tesla orders and attract other clients.

Source: Samsung to Begin EUV Trials at Taylor Fab in March, Make Chips for Tesla 

At its Vision 2025 conference, Intel revealed it has started risk production of its advanced 18A process node. This is a pivotal step marking the beginning of new low-volume test manufacturing for this technology.  

Intel’s Kevin O’Buckley, Senior Vice President of Foundry Services, made the announcement. Intel nears full completion of its five nodes in four years. This initiative was launched by former CEO Pat Gelsinger as part of Intel’s quest to retake the semiconductor crown from TSMC. The conference also marks the first time new CEO Lip-Bu Tan has taken to the stage as Intel’s leader.  

Launched in 2021, Intel’s four-year 5N4Y plan shifted focus. The company canceled high-volume 20A production due to cost. However, the 18A node is nearing completion. The plan aims to make process nodes available for production, not necessarily immediate high-volume manufacturing.  

Risk production is a key step toward launching a new process node, demonstrating Intel’s confidence that the node is nearly ready for HVM. Leading up to this stage, the company has already produced numerous 18A test chips and shutters, often prototyping multiple designs on a single wafer.  

During risk production, Intel manufactures wafers with a single chip design in low volumes. The company updates its manufacturing process and qualifies the Node and Process Design Kit (PDK) in real-life runs. Production will scale up in the second half of the year. This stage follows R&D, Design, and Prototyping.  

Risk production involves some uncertainty for customers. Yields and functionality may fall short of targets, while manufacturing techniques and tooling are optimized. During this period, customers typically produce qualification or engineering samples using the new process. These early chips may not have guaranteed yields. However, they enable customers to begin product validation and prepare for full-scale launch when high-volume manufacturing is achieved.  

Nonetheless, some customers choose to accept these risks to gain early access to the node, which enables them to improve their designs and achieve time-to-market advantages over competitors.  

Intel has not specified whether 18A risk production is for its Panther Lake processors expected later this year or for external foundry customers. However, Panther Lake will enter mass production this year and is likely the focus of risk production. The timeline aligns with Intel’s typical risk-based production to HVM schedules.  

While Intel introduced several new technologies with its canceled 20A mode, the 18A (1.8nm) chips will be the first to feature both PowerVia backside power delivery and ribbon FET gate all-around (GAA) transistors. PowerVia improves power routing performance and transistor density. Ribbon FET enhances transistor density and switching speed within a smaller area.  

Intel is also advancing its wider foundry roadmap, which includes the upcoming 14A node. It’s the first to use high NA EUV lithography. Additional node extensions will expand Intel’s foundry services portfolio to serve a wider range of applications.  

These developments come as Intel Foundry navigates changing macroeconomic conditions. For example, while Intel has delayed its Ohio expansion until 2030, the 18A risk production announcement aligns with positive reports on initial 18A wafer runs in Arizona, reinforcing the company’s adaptability.  

Industry observers anticipate further details about Intel’s future plans at the Foundry Direct Connect event in late April, which promises to provide additional context for Intel’s current risk production efforts.  

Risk production, while it sounds scary, is actually an industry-standard terminology. The importance of risk production is that we’ve gotten the technology to a point where we’re freezing it, O’Buckley explained that our customers have validated that 18A is good enough for my product, and we now have to do the risk part, which is to scale farm, making hundreds of units per day to thousands, tens of thousands, and then hundreds of thousands. Risk production is scaling our manufacturing and ensuring we can meet not just the technology’s capabilities but also those at scale.

Source: Intel announces 18A process node has entered risk production

News Summary 

• AMD introduces Ryzen AI 400 and Pro 400 series processors. The Ryzen AI 400 series targets consumer and commercial devices with up to 60 NPU TOPS for Co-Pilot+ PCs and AI features. The Pro 400 series is aimed at business users seeking enhanced manageability and security.  

• AMD introduces new Ryzen AI Max+ SKUs, bringing high-performance AI and graphics to ultra-thin notebooks, workstations, and small-form-factor devices for creation, gaming, and AI development.  

• AMD unveils the Ryzen AI Halo A Mini PC, delivering Ryzen AI Max+ performance for AI developers and offering an out-of-the-box experience that accelerates AI innovation at the edge.  

• AMD announced the Ryzen 7 1950X3D, its flagship gaming processor based on Zen 5 architecture with AMD 3D V-Cache, designed for enthusiasts prioritizing top-tier gaming performance. This model stands apart from AI-focused SKUs.  

• AMD sees strong year-on-year growth in OEM adoption of Ryzen AI processors, with more systems launching across consumer, commercial, and gaming segments throughout 2026.  

• AMD announces AMD ROCm 7.2 software for Windows and Linux, bringing seamless support for Ryzen AI 400 series processors and inclusion in ComfUI.  

At CES 2026, AMD also revealed its latest mobile and desktop processors, expanding its client computing portfolio. This launch underscores AMD’s drive to lead in AI capabilities, premium gaming performance, and commercial-ready features, bringing these advances to more systems and users than ever before.  

AMD introduced the new AMD Ryzen AI400 series for Co-Pilot Plus PCs and Ryzen AI Max+ processors for premium ultra-thin and light notebooks and small-form-factor desktops. The company also announced the Ryzen AI Pro 400 series, enabling AI acceleration, modern security, and enterprise-class manageability to meet the needs of today’s business landscape. Recognizing AI as central to the PC experience, AMD is strengthening its hardware portfolio with AMD Ryzen AI Halo, the company’s first branded AI developer platform. AMD pairs this hardware with new ROCm 7.2 software support for Ryzen AI 400 series processors and an AI bundle for AMD Software Adrenaline Edition, ensuring AI adoption and development are more accessible than ever.  

AMD announces the Ryzen 7 9850X3D, an improved gaming CPU with a higher boost clock built on Zen 5 and 3D V-Cache. Radeon users get FSR Redstone for ML frame generation and upscaling in new AAA games.  

The PC is being redefined by AI, and AMD is leading that transformation, said Jack Huynh, senior vice president and general manager of the AMD Computing and Graphics Group. Across consumer, commercial, and enthusiast teams’ systems. We are delivering platforms that bring high-performance computing, leadership AI, interactive graphics, and a growing software ecosystem that strengthens developers and creators, so intelligence is built in, performance and effectiveness scale smoothly, and innovation reaches every form factor. Our full-stack approach is coming to life, enabling a smarter, faster, and more absorbing experience for users today and tomorrow.  

AMD ROCm Software Experience Developer Access 

AMD announced that AMD ROCm, the open software platform, now supports Ryzen AI 400 series processors and is available for download through Confi UI. The upcoming AMD ROCm software 7.2 release will extend compatibility throughout both Windows and Linux, and new PyTorch builds can now be easily accessed through AMD software for simplified deployment on Windows.  

Over the past year, AMD ROCm software has delivered up to a five times improvement in AI performance. Platform support has doubled across Ryzen and Radeon products in 2025, and availability now spans Windows and additional Linux distributions, resulting in a year-on-year increase of up to 10x in downloads. Together, these updates make AMD ROCm software a more powerful and accessible foundation for AI development, reinforcing AMD as a platform of choice for developers to build the next generation of intelligent applications. 

Source: AMD Expands AI Leadership Across Client, Graphics, and Software with New Ryzen, Ryzen AI, and AMD ROCm Announcements at CES 2026 

Built for reliable AI in production, GPT 5.4 offers stronger reasoning, dependable execution, and scalable agent workflows.  

We are excited to share that OpenAI’s GPT-5.4 is now available in Microsoft Foundry. This model helps organizations move from planning to reliably completing work in real production settings. As AI agents handle longer, more complex workflows, consistency and follow-through are just as important as they entail. GPT-5.4 offers stronger reasoning and built-in computer-use features to support automation and reliable execution across tools, files, and multi-step workflows.  

GPT-5.4 Enhanced Dependability in Production AI 

GPT-5.4 is designed for organizations running AI in production, where consistency, instruction-following, and context retention are crucial. It advances reasoning, coding, and agent workflows to help AI not just plan but complete tasks with fewer interruptions and less supervision.  

GPT-5.4 is more stable during extended interactions than earlier versions, giving teams confidence to depend on agent-based AI for daily needs.  

GPT-5.4 brings new capabilities created for production-grade AI:  

• It provides more consistent reasoning, maintaining clear intent across complex, multi-step interactions.  

• Instruction alignment is improved, with less need for tuning and oversight.  

• Performance is faster, making workflows more responsive for real-time use.  

• It includes built-in computer-use features for organizing tools, accessing files, extracting data, running code safely, and handling tasks between agents.  

• Tool use is more reliable, reducing the need for prompt tuning and human monitoring.  

• It generates higher-quality outputs like documents, spreadsheets, and presentations with a more consistent structure.  

These improvements ensure more predictable AI performance for longer, complex tasks.  

Turning Capabilities Into Actual Results 

GPT-5.4 delivers practical value in production, where reliable task completion is critical:  

• Agent-driven workflows such as customer support, research assistance, and business process automation  

• Enterprise Managed Work, including drafting documents, analyzing data, and generating presentation-ready outputs.  

• Developer workflows spanning code generation, refactoring, debugging, support, and UI scaffolding  

• Extended thinking tasks where logical uniformity must be preserved across longer interactions  

Teams using GPT 5.4 in production experience less task drift, fewer workflow failures, and improved output predictability.  

GPT 5.4 Pro: Deeper Analysis for Complex Decisions 

GPT 5.4 is an enhanced version designed for situations where deep, comprehensive analysis is critical, and it is optimized for reliable execution and follow-through in production tasks.  

Additional capabilities include:  

• Multipath reasoning evaluation allows alternative approaches to be explored before selecting a final response.  

• Greater cognitive depth supporting problems with trade-offs or multiple valid solutions  

• Its improved stability supports sustained analytical tasks along long reasoning chains.  

• It offers enhanced decision support when thoroughness outweighs speed.  

Organizations select GPT-5.4 Pro for thorough analysis of complex challenges, such as scientific research, while GPT-5.4 is ideal for reliable task execution with strong follow-through.  

Microsoft Foundry: Enterprise Grid Control From The Start 

Organizations access GPT-4 and GPT-5.4 via Microsoft Foundry, which provides controls for Responsible Production-Grade AI. Foundry simplifies policy enforcement, monitoring, versioning, and auditability, helping teams manage AI over time.  

With GPT-5.4 in Microsoft Foundry, organizations can add advanced agent features to existing systems and meet security, compliance, and operational requirements from the start. 

Source: Introducing GPT-5.4 in Microsoft Foundry 

AWS has launched a new, stateful runtime environment for AI agents in Amazon Bedrock, developed in collaboration with OpenAI. This environment supports long-running, complex AI workflows that keep context and memory across multiple steps and sessions.  

Key Features and Benefits 

This new stateful runtime marks a major change from traditional stateless AI runtimes, which handle each request separately.  

• Persistent context and memory: agents maintain a consistent context, including conversation history, tool state, and identity boundaries. Developers no longer need to build external state management systems.  

• Long Running Tasks: The runtime manages complex, asynchronous workloads and multi-step processes for hours or days. Agents work independently on projects like customer support, IT automation, or data processing.  

• Simplified production deployment: teams focus on business logic as orchestration and state are fully managed by the new architecture.  

• Integration with the AWS Ecosystem: This environment operates within the customer’s AWS environment and integrates with AWS security tools and identity systems, such as Amazon Cognito, as well as governance tools.  

• Powered by OpenAI Models: The Runtime Environment incorporates models from OpenAI that have been specifically optimized for AWS and are offered through Amazon Bedrock as part of the AWS OpenAI partnership. Microsoft Azure will still be the only provider of stateless OpenAI APIs.  

Impact on Enterprise AI 

This AWS OpenAI partnership indicates a shift toward agentic infrastructure as a platform for businesses. This means:  

• Accelerated development: It reduces the time to production-ready AI agents from months to weeks.  

• Improved capabilities: It allows for more advanced applications, such as customer support across multiple systems, sales operations workflows, and internal IT automation with approvals and audits  

• Governance and control: The architecture provides a managed pipeline using an AI-driven lifecycle (AI-DLC) framework that assesses agents for performance, cost, and security before deployment.  

The stateful runtime environment should be available through Amazon Bedrock in the next few months.  

AI agents are great at reasoning, but the real challenge is making sure they can reliably handle multi-step tasks over time using real tools and systems with proper controls.  

Today, we are making this easier by partnering with Amazon to launch a new stateful runtime environment that runs directly on Amazon Bedrock. AWS customers can use this environment, powered by OpenAI models and optimized for AWS, to support agent workflows with the state reliability and governance needed for production.  

Making It Easier to Bring Agents Into Production 

Many Agent prototypes based on stateless APIs tackle simple use cases:  

• One prompt  

• One answer  

• Maybe one call to a tool.  

Production work is different. Real workflows unfold across many steps. They require context based on previous actions. They depend on multiple tool outputs, approvals, and system state, and need trusted guardrails to secure environments.  

With stateless APIs, development teams must build the orchestration layer themselves. They need to decide how to store state, call tools, handle errors, and safely resume long-running tasks.  

The Stateful Runtime Environment is built to make this easier. It runs within your AWS environment and works well with AWS services. Instead of assembling separate requests, your agents can now automatically handle complex tasks with context-carrying forward memory, workflow state, environment, use, and permission boundaries.  

What Can You Do With This 

Now it’s easier to build solutions like:  

• multi-system customer support  

• sales operational workflows  

• internal IT automation  

• financial processes that include approvals and audits  

Faster Time to Production for Multi-Step Workflows 

When the runtime manages orchestration and state across steps, teams can focus on the workflow and business logic rather than building additional support systems.  

Designed For Long-Running Tasks 

Stateful tasks are built to run reliably over time, keeping the context and control boundaries needed for multi-step work.  

AWS Native Deployment and Governance 

To understand how this stateful runtime can benefit your organization, reach out to your OpenAI team or request a contact from us today.

Source: Introducing the Stateful Runtime Environment for Agents in Amazon Bedrock 

NVIDIA recently released the GeForce Game Ready driver 595.59 to improve performance in Resident Evil: Requiem, but, according to a machine translation, this update has caused problems with RTX 3000-series and newer cards. Users report that the driver only detects one fan on their GPUs.  

Some people suspected that third-party apps like MSI AfterBurner were causing the issue; however, another user experienced the same problem even without AfterBurner installed.  

NVIDIA appears to have removed the driver update, as it is no longer on their website (a driver is the software that allows the operating system to communicate with your graphics card). If you have already installed the latest driver and are experiencing problems, you should roll back to the previous version. To do this in the NVIDIA app, click the three dots in the Drivers tab.  

If you do not have NVIDIA’s software open, Windows Device Manager, expand Display Adapters, and double-click your GPU in the Properties window. Go to the Drivers tab and select Roll Back Driver. If that option is unavailable, NVIDIA recommends uninstalling the GPU driver and reinstalling the latest available version to solve the issue, since the problematic driver has been removed.  

The new NVIDIA 595.71 driver has brought new problems not present in last week’s troubled 595.59 release, which NVIDIA pulled a few days ago. As a result, several users and at least one YouTuber have found that the 595.71 driver limits GPU overclocking on many RTX 40 and 50 series cards. The most affected models lose about 200 MHz of overclocking headroom compared to earlier drivers.  

The issue seems to be artificial voltage limits added to the in driver 595.71, either by mistake or on purpose. YouTuber bang4buckpcgamer showed that his Asus TUF Gaming RTX 5090 lost 65 mV of voltage headroom, keeping the card below 1 W. This change reduced his overclocking headroom by about 171 MHz, dropping from 3165 MHz to just under 3000 MHz. This only happens when the offset is about 150 MHz. With a 150 MHz offset or less, the GPU does not restrict voltage and can reach up to 1.060 V. Similar issues have also been shared on the NVIDIA forums. One user with an RTX 5080 reported that their GPU used to hit 3,100-3,200 MHz with previous drivers and can now only reach 2,395 MHz with 595.71.80 owner published their own 3DMark scores with the previous 591.86 driver compared to 591.71 with a hefty 450 MHz GPU overclock. They found the new driver was running the GPU 300 lower and pulling 43 fewer watts than from 403W to 360W.  

However, not all RTX 50 series GPUs appear to be affected. Curiously, three commentators on bang4buckpcgamer’s aforementioned YouTube video with a Gigabyte Aorus Master RTX 5090 graphics card report having no restrictions whatsoever. Another RTX 5090 owner with a PNY EPYC OC variant reported no issues achieving a max overclock of 3157 MHz with the latest driver. Two RTX 5070 owners, one with an Asus variant and the AMD MSI Gaming Trio OC, also reported no issues.  

This variability suggests that some owners may simply be lucky with their hardware, so their cards’ voltage and frequency scaling are not affected by this bug. Still, the new issue has led to many angry comments from gamers, with some blaming AI code for causing problems with Nvidia’s drivers. NVIDIA has officially not recognized the issue yet, but the artificial voltage limits do appear to be a bug rather than an official change. The NVIDIA patch notes don’t mention any new voltage limits, and certain GeForce RTX GPU models apparently aren’t subject to any limitations when running 595.71. We’ll have to see whether the company issues another corrective release or a fixed driver in the near future, along with any further explanation of the issue.

Source: Nvidia rolls back its latest driver update — Game Ready Driver 595.59 reportedly causes fan issues on RTX 3000, 4000, and 5000-series GPUs

Last week, Apple released the first iOS 16.4 beta, which included a bunch of new features and changes to Apple’s mobile software. Included was the news that Apple had started testing encrypted RCS messaging, but only between iPhones. Now, with the release of iOS 16.4 beta 2, that’s changed.  

At the time, I hoped Apple would soon test encryption for RCS messages between iPhones and Androids. I did not expect this to happen quickly, but I am glad it has.  

iPhone beta testers must have the latest iOS 16.4 beta installed to participate in cross-platform testing. Android users will need to have the latest version of Google Messages.  

Apple has affirmed that this feature will be in testing for a while. It will not ship the final version of iOS 26.4 and is not available for all devices and carriers. You have to be one of the few beta testers to send encrypted messages to your Android-using friends.  

Apple has previously confirmed that RCS won’t change the green bubble situation, so Android users will remain green regardless of encryption status, while iMessages are displayed in blue. However, testers will see a lock icon on all encrypted messages, indicating the security of their conversations.  

That change applies to our CS and iMessage, so there is absolutely no confusion. No lock means your messages are about as secure as an open gate.  

The benefit of RCS messaging between iPhone and Android is that all cross-platform issues are gone. Larger file-sharing limits mean photos and videos are not heavily compressed, and users get modern features like real-time typing indicators and reaction emojis.  

Soon, everyone will benefit from end-to-end encryption, which secures messages from anyone trying to intercept them. We do not know when it will be available to all. Apple has only said it will arrive in future releases of iOS, iPadOS, macOS, and watchOS 26.  

Ultimately, we will have to wait until Apple confirms everything works as intended.

Source: iOS 26.4 beta 2 now lets iPhones send encrypted RCS messages to Android — here’s how it works 

Meta engineers have launched KernelAgent, a multi-agent system that automates the creation and tuning of GPU kernels for AI workloads. This open-source tool, available in the Meta-pyTorch/KernelAgent GitHub repository, uses large language models and a hardware-guided feedback loop to generate fast, verifiable Triton kernels from PyTorch programs.  

Key Features 

• Multi-agent system: KernelAgent splits the complex task of kernel optimization into dedicated roles. ProfilerAgent monitors and collects hardware performance data; JudgeAgent analyzes requests to identify areas for improvement; and the Optimization manager coordinates the workflow and decides which optimizations to pursue. These agents work together in cycles.  

• Hardware-guided optimization: Instead of relying on static models like traditional compilers, KernelAgent bases its choices on real hardware performance data, including compute usage and memory bandwidth, gathered with NVIDIA Nsight Compute (NCU).  

• Ongoing feedback loop: The system uses a closed-loop workflow.  

1. Profiling: The system collects hardware metrics when it runs the kernel for the first time  

2. Diagnosis: A powerful language model reviews the data to find performance bottlenecks.  

3. Optimization: another large language model creates an improved kernel based on these suggestions.  

4. Verification & Benchmarking: The system tests the new kernel to ensure it is accurate and performs well.  

5. Iteration: The process repeats, and agents learn from previous successes and failures saved in shared memory.  

• The Optimization Manager explores several optimization paths in parallel, keeping only the best-performing kernels.  

• KernelAgent identifies and fuses parts of PyTorch programs, replacing them with optimized Triton kernels.  

Performance 

On 100 L1 KernelBench tasks, KernelAgent achieved 2.02x speedup over previous kernels and averaged 1.56x faster than the default torch. compile, reaching 89% of hardware efficiency on an H100 GPU.  

Optimizing GPU kernels is becoming more important for today’s AI workloads. As models get bigger and more specialized, performance increasingly depends on kernel efficiency rather than just the algorithms. However, manually tuning kernels requires significant expertise and an in-depth understanding of GPU hardware, memory, and performance trade-offs. The challenge only grows as more channels and kernels are added, and each new GPU architecture requires new optimization strategies.  

In practice, skilled kernel engineers use a step-by-step approach to optimize kernels. They profile kernels with tools such as NVIDIA Nsight Compute and examine hardware performance counters to identify bottlenecks and make targeted improvements.  

They ask questions like:  

• Is the tiling strategy missing out on memory bandwidth?  

• Does the kernel need a full redesign rather than just parameter tuning?  

Often, they have to evaluate several kernel designs, each with its own bottleneck, before finding one that fully utilizes the hardware. This process works well, but it usually takes days or even weeks.  

Modern compiler stacks have made big advances in automating kernel generation. For example, Torch.compile captures computation graphs and generates Triton kernels via graph transformations. Pattern matching and compiler rules, as well as other systems like TVM and XLA, employ similar tools to handle many common kernel patterns and deliver good performance from the start. Still, most compiler rules rely on static models rather than real measurements from running or actual hardware.  

KernelAgent seeks to automate this diagnosis-driven optimization process by harnessing real hardware signals to steer Kernel’s tuning of forward-pass (inference) kernels, which directly impact latency and throughput. This system is built on three fundamental principles:  

• For every hardware decision, both bottleneck identification and optimization selection should be based on precise profiling data.  

• Adapt multiple optimization tactics concurrently. Given identical hardware data, there may be several viable optimization pathways. KernelAgent evaluates these alternatives in parallel, saving time and synthesizing previous strategies to generate superior algorithms.  

• Iterate by learning from every round using shared memory. Optimization agents review what succeeded or failed, storing insights collectively to inform future cycles and avoid repeating errors. 

Source: KernelAgent: Hardware-Guided GPU Kernel Optimization via Multi-Agent Orchestration 

The Gemini app provides embedding models that generate embeddings for text, images, video, and other content types. You can use these embeddings for activities such as semantic search, classification, and clustering, which often yield more accurate, context-aware results than keyword searches.  

The newest model, Gemini-Embedding-2-Preview, is the first from Gemini API to handle multiple content types, mapping text, images, video, audio, and documents into one shared embedding space. This enables searching, classification, and clustering across over 100 languages. For more details, check out the Multimodal Embedding section. If you only need text, Gemini-Embedding-001 remains available.  

If your product relies on retrieval, augmented generation (RAG) embeddings are crucial for making these systems more accurate, coherent, and context-aware for teams seeking a managed RAG solution. A file search tool makes RAG management easier and more affordable.  

Google has launched Gemini Embedding 2 for public previews, bringing enhancements over the previous version.  

As Google’s first native multimodal embedding model, Gemini Embedding 2 can map text, images, video, and documents into one shared embedding space. It was released alongside new AI features for Workspace apps.  

If you are new to this, embedding models are different from generative models like Gemini 3. Embedding models help computers understand context by turning text, images, or video into vectors, which are mathematical formats that computers can read and analyze. These embeddings yield more context-aware results across tasks such as semantic search, classification, and clustering than keyword-based methods.  

The first Google Embedding model only worked with text. Gemini Embedding 2 now supports text, images, videos, audio, and documents in a single unified embedding space across 100 languages. Below are the content limits:  

• Text: up to 8192 tokens per request.  

• Images: up to 6 images per request, supporting PNG and JPEG formats.  

• Video: up to 120 seconds of video in MP4 or MOV format per request.  

• Audio: processes and embeds audio data directly without needing transcriptions.  

• Documents: can be PDFs up to 6 pages long.  

In a blog post, Google said the new model streamlines complex pipelines and enhances a wide variety of multi-modal downstream tasks from retrieval-augmented generation (RAG) and semantic search to sentiment analysis and data clustering. The model can analyze detailed relationships among different media types by accepting multiple media types in a single request, such as images and text.  

For example, Google noted that Gemini embeddings can help legal professionals find important information during the discovery phase of litigation. The multimodal embedding improves precision and recall across millions of records and enhances image and video search.  

Gemini embeddings (Gemini-embeddin-2-preview) are now available for public preview through Gemini, the Gemini API, and Vertex. The Gemini-embedding-001 model is still available for text-only needs.

Source: Embeddings 

Google releases Gemini Embedding 2 AI model with multimodal support