XTHE

OpenAI released the O3 Mini in early 2025. This smaller reasoning model is built for STEM tasks and offers better intelligence and lower latency than the O1 Mini. Users can choose low, medium, or high reasoning effort. It can also be used to create high-value, cost-effective, real-time use.  

STEM Raising Performance 

O3 Mini performs very well at a medium effort level. It equals the larger O1 model’s functions but responds faster and more accurately.  

• In the 2024 AIME Math Test, O3 Mini with high reasoning effort did better than both O1 Mini and the full O1 model.  

• Reputation coding (Codeforces): O3 Mini achieves an ELO rating of 27/27, significantly higher than O1 Mini’s 1891. It outperforms the O1 Mini high in programming tasks.  

• On the GPQA Diamond Science Test, O3 Mini scored 87.7% on PhD-level questions, which is 10% better than O1 Mini.  

• In software engineering tests (SWE Bench Verified), O3 Mini is the top performer in its series. With significant effort, it solves complex tasks and often beats the O1 Mini by more than 20% in software benchmarks.  

• Experts found that O3 Mini made 39% fewer major errors than O1 Mini.  

Latency and Speed Metrics 

• Training effort customization: users can select low, medium, or high effort to trade off speed and accuracy. Models are 63% cheaper than the O1 Mini.  

• Advanced features:  

• Supports function calling  

• Structured outputs and developer messages  

• One limitation is that O3 Mini does not support vision features like the O1 model.  

Comparison with other models. 

• Compared to the O1 Mini, the O3 Mini is faster, cheaper, and more accurate. It also does better in coding and math competitions.  

• Versus DeepSeek R1: While DeepSeek R1 is frequently more cost-effective per token, O3 Mini is generally faster for live coding and STEM tasks, and it shows superior safety with a lower rate of unsafe responses (1.19% versus 11.98%).  

Today we’re launching OpenAI O3 Mini, our latest and most affordable reasoning model, now available in ChatGPT and through the API. First reviewed in December 2024, this fast and capable model pushes the limits of small models, offering strong STEM skills, especially in science, math, and coding. It also keeps the low cost and fast response times of OpenAI O1 Mini.  

OpenAI O3 Mini is our first small reasoning model to support popular developer features like function calling, structured outputs, and developer messages, making it ready for production use right away. Like OpenAI O1 Mini and O1 Preview, O3 Mini also supports streaming. Developers can choose from three reasoning effort options: low, medium, or high, to fit their needs best. This means O3 Mini can focus more on tough problems or work faster when speed matters. O3 Mini does not handle vision tasks, so developers should use OpenAI for those starting today. O3 Mini is available in the Chat Completions API, Assistance API, and Batch API for select developers in API user tiers 3-5.  

Starting today, ChatGPT Plus, Team, and Pro users can use OpenAI O3 Mini, and Enterprise users will get access in February. O3 Mini will replace O1 Mini in the model picker, offering higher rate limits and faster responses. This makes it a great choice for coding, STEM, and logic tasks. Plus and Team users will now have their daily message limit increased from 50 to 150 with O3 Mini. O3 Mini also now supports search, helping users find current answers with links to web sources. The search feature is an early prototype while we work to add it to more models.  

Free plan users can now try O3 OpenAI O3 Mini by choosing a region in the message composer or by regenerating a response. This is the first time a reasoning model has been available to two free ChatGPT users.  

OpenAI O1 remains our main model for general knowledge, while OpenAI O3 Mini is designed for technical fields that require accuracy and speed. In ChatGPT, O3 Mini uses a medium level of understanding effort to balance speed and accuracy. Paid users can also choose O3 Mini High in the model picker for a smarter version that takes a bit longer to reply. Pro users get unlimited access to both O3 Mini and O3 Mini High.  

Fast, Powerful, And Built For STEM Reasoning 

Tech: OpenAI O1 and O3 Mini are tuned for STEM reasoning. With medium reasoning effort, O3 Mini matches O1’s performance in math, coding, and science, but responds faster. Expert testers found that O3 Mini provides more accurate, clearer answers with better reasoning than O1 Mini. They preferred O3 Mini’s answers 56% of the time and saw 39% fewer major errors on tough real-world questions. With moderate effort, O3 Mini matches O1’s results on challenging tasks, such as AIME and GPQA.  

What’s Next? 

The launch of OpenAI O3 Mini is another step forward in our effort to make cost-effective intelligence possible. We have improved the rationale for STEM fields and kept costs low so people can access high-quality AI. Since GPT-4, we have reduced per-token pricing by 95% while still offering strong reasoning abilities. As more people use AI, we are committed to leading the way by building models that are smart, efficient, and safe at scale.

Source: OpenAI o3‑mini 

Decompression helps reduce storage costs and speed up data transfers across databases, data centers, high-performance computing, deep learning, and other areas. However, decompressing this data can slow things down by adding latency and using valuable computing power.  

To address these challenges, NVIDIA introduced the hardware decompression engine DE in the NVIDIA Blackwell architecture and created it for the nvCOMP library. Together they offload decompression from general-purpose compute, accelerate widely used formats like Snappy, and make adoption seamless.  

In this blog, we will explain how DE and nvCOMP work, share usage tips, and highlight the performance benefits they deliver for data-intensive tasks.  

How The Decompression Engine Works 

The new DE in the Blackwell architecture is a dedicated hardware block that speeds up decompression for snappy LZ4 and D-flat-based streams. By handling decompression in hardware, the DE lets streaming multiprocessor (SM) resources focus on computation rather than data movement.  

The DEE is built into the copy engine. You no longer need to do host-to-device copies and then run software decompression. Now, compressed data can move directly over PCIe or C2C and be decompressed as it travels, helping remove a major I/O bottleneck.  

The DE does more than boost throughput. It allows data movement and computation to occur simultaneously. With multi-stream workloads, decompression can run in parallel with SM kernels, so the GPU stays busy. This helps with data-intensive tasks like training LLMs, evaluating large geonomics datasets, and/or running HPC simulations, keeping up with the high bandwidth of Blackwell GPUs without being slowed down by I/O.  

The Benefits of NvComp’s GPU-Accelerated Decompression 

The NVIDIA nvCOMP library offers GPU-accelerated routines for both compression and decompression. It works with many standard formats, as well as formats that NVIDIA has tuned for top GPU performance.  

Standard formats, CPUs, and fixed-function hardware often have an edge over GPUs because GPUs have less parallelism for these tasks. The decompress engine solves this issue for many workloads. Next, we will explain how to use nvCOMP with the DE.  

How to use DE and nvCOMP 

Developers should use DE through the nvCOMP API’s. Right now, DE is only on certain GPUs (B200, B300, GB200, and GB300), so using nvCOMP lets you write code that works among different GPUs as support grows. If DE is available, nvCOMP uses it automatically. If not, it switches to its fast SM-based methods without needing changes to your code.  

To ensure this works on DE-enabled GPUs, follow these steps. nvCOMP usually accepts any input and output buffers that the device can access, but the DE has stricter rules. If your buffers don’t meet these rules, nvCOMP will use the SM or decompression instead.  

You can use cudaMalloc as usual for device-to-device decompression. For host-to-device or host-to-host decompression, use cudaMallocFromPoolsync or active cuMemCreate, but make sure to set up the allocators correctly.  

How SM Performance Compares to DE 

DE offers faster decompression and lets the SM handle other tasks. The DE has dozens of execution units, while SMs have thousands of volts. Each DE unit is much faster at decompression, but in some cases, a fully loaded SM can come close to DE speed. Both SM and DE can use cost-pinned data as input, enabling zero-copy decompression.  

The next figure shows how SM and DE perform on the Silesia benchmark for LZ4, D-flat, and Snappy algorithms. Snappy has been newly optimized in nvCOMP 5.0, and there are more chances to improve D-flat and LZ4 as well.  

Performance was measured using 64KiB and 512KiB chunk sizes on both small and large data sets. The large data set is the full Silesia dataset, and the small data set is the first 50 MB of Silesia.tar.  

Get started 

The Decompressor engine in Blackwell helps solve one of the biggest problems in data-heavy workloads: getting first efficient decompression. By moving this job to a dedicated hardware application, it runs faster and frees up GPU resources. For other tasks, operators can take advantage of these improvements without changing their code, leading to better pipelines and better performance.

Source: Speeding Up Data Decompression with nvCOMP and the NVIDIA Blackwell Decompression Engine 

Samsung Electronics has announced the One UI 8.5 Beta program, which brings simpler ways to create, connect, and stay secure. With this update, CUV users can get more done with less effort, thanks to easier actions, easier device management, and better security.  

One UI 8.5: What’s New for Content Creation 

One UI 8.5 makes it easier to create and share content. With the updated photo assist, users can keep making new images without stopping. They can edit photos as much as they want, without saving each step along the way, when they are done. It is easy to review the edit history and choose their favorite versions.  

Sharing is now easier with improvements to Quick Share, which can recognize people in photos and suggest sending the images directly to those contacts.  

How does One UI 8.5 improve device connectivity? 

New cross-device feature: make it simpler to manage devices, share files, or connect with nearby devices. Audio broadcast lets users easily communicate with nearby LE audio-supported devices via AuraCast. Now users can also broadcast their voices through their Galaxy phones’ microphones, which is useful for group activities such as tours and events.  

Storage Share connects the Galaxy ecosystem by letting users view files from other Galaxy devices, such as tablets or PCs, right in the My Files app. It also lets users access their phone’s files from other Samsung devices, including their TV.  

How Is My Galaxy Device Staying Protected 

One UI 8.5 enhances device security and makes security settings easier to manage. Thief protection helps keep phones and data safe. If a device is lost or stolen, the failed authentication lock will lock the screen after too many failed attempts to unlock it with a fingerprint, PIN, or password. The identity check now protects more settings, adding extra security.  

Availability 

The One UI 8.5 beta program will first be available for Galaxy S25 series users in select markets, including Germany, India, Korea, Poland, the UK, and the US. Starting December 8, Galaxy users can sign up for the beta through the Samsung Members app.  

• Use the Generative Edit feature in Photo Assist. You need a network connection and a Samsung account. When you edit with Generative Edit, your photo may be resized. A visible watermark will appear on the saved image to indicate that the AI created it. The screen image shown is for illustration only. The actual user experience and interface may be different.  

• This feature is available on devices running One UI 2.1 or later with Android Q or newer. Quick Share needs both Bluetooth Low Energy and Wi-Fi to work. The actual speed can change based on your device, network, and environment. The screen image is for illustration only. The real user experience and interface may look different.  

• This feature is only available on Galaxy S25 series devices, which include the Galaxy S25, S25 Plus, and S25 Ultra.   

• Each device must be signed in to the same Samsung account and must have both WiFi and Bluetooth turned on. Requires updated Galaxy phones and tablets with One UI 7 or higher, kernel version 5.15 or higher, and Galaxy Book 2 or later (Intel and Galaxy Book 4 or later) (Arm) and Samson Smart TV models, including U8000 and above, released after 2025. Feature availability may vary by region and device model. Screen image simulated for illustrative purposes; actual UX/UI may vary. 

Source: Samsung Launches One UI 8.5 Beta for Next-Level Ease of Use 

We are pleased to share that our Surface Co-Pilot Plus PC lineup for business is growing. The new 13.8-inch Surface Laptop 5G with Intel Core Ultra (Series 2) processors will start shipping on August 26th. You can get the new 13-inch Surface Laptop and 12-inch Surface Pro starting today.  

AI gives organizations a real advantage, but only if it’s available when needed. The Surface Laptop 5G delivers this with a neural processing unit (NPU) that handles over 40 trillion operations per second, enabling everyday tasks to run faster and easier. Whether you need to stay focused in meetings, find information quickly, or cut down on routine work, AI helps you get more done. With built-in 5G, you can stay connected to Microsoft 365, Co-Pilot, and other cloud tools for better insights and real-time teamwork.  

Many business customers have asked for 5G in the Surface Laptop. This request concerned more than just adding a modem. It showed a real need for instant, secure, and reliable connections, without worrying about signal strength or finding a hotspot.  

This need inspired the design of the Surface Laptop 5G. Our goal was to create the best 5G-connected laptop, where being connected is effortless. Whether you are a consultant or a team’s call on a train, a field engineer, uploading site data, or a sales leader finishing a proposal in a hotel lobby, Surface Laptop 5G helps you keep working wherever you are.  

Surface Laptop shows how Microsoft brings together hardware, software, and cloud services to create smart, secure, and connected tools for today’s mobile workforce. The Surface for Business line-up has solutions for every need, from flexible tablets to powerful laptops, all backed by Microsoft’s management tools and top security.  

Engineered For Seamless 5G Performance 

Adding 5G to the Surface Laptop took more than just adding a modem. We redesigned every part of the Surface Laptop 5G to ensure it’s seamless, reliable, and secure while maintaining the design, performance, and portability you expect.  

Dynamic Antenna System 

At the heart of Surface Laptop 5G is a dynamic antenna system that constantly adapts to its environment. With six strategically placed antennas, the device automatically adjusts signal paths and power based on how it’s being held or used, ensuring strong, reliable connectivity exactly when and where it’s needed. As users move between environments, the device smoothly transitions between 5G and Wi-Fi networks, maintaining a steady, secure connection to cloud-based apps, updates, and corporate resources. This novel antenna design also enables Surface Laptop 5G to act as a mobile hotspot, securely sharing its 5G connection with other devices when Wi-Fi isn’t available.  

Carefully Designed Hardware 

Most laptops put the antenna near the base, where signals can be blocked by objects on your legs. The Surface Laptop 5G is different. Its antennas are placed higher up to reduce interference and keep your connection strong and steady.  

To make this design work, we needed a new material that lets radio signals pass through while remaining durable, high-quality, and light. We created a custom, multi-layered laminate that delivers all this, so you get reliable 5G connectivity without sacrificing portability or style.  

Surface Laptop 5G also includes both Nano SIM and eSIM options integrated to preserve its slim profile while enabling worldwide connectivity. Weighing under 3 lb, it is light and easy to carry across campuses, through airports, or from meeting to meeting.  

Tested For The Real World 

In Surface, we don’t just emulate real-world use; we build for it. The Surface Laptop 5G was tested in homes, apartments, and active work environments to replicate typical scenarios, including:  

• moving between rooms  

• switching networks  

• multitasking  

• working from laptops  

These helped us fine-tune antenna placement, thermal performance, and connectivity behavior to reflect how people actually work.  

To ensure it works everywhere, we tested the 5G hardware with over 100 mobile operators across more than 50 countries. This means you get reliable 5G connectivity wherever your team goes.  

Secure and Connected for Smarter Management 

With built-in 5G, Surface laptop devices are persistently connected, so IT can send security updates and force policies, and get instant insights from almost anywhere. This is possible because Surface works closely with Windows and InTune, bringing together hardware, software, and cloud management into a single Microsoft solution. Surface devices paired with Windows AutoPilot enable a true zero-touch deployment experience where devices arrive carefully configured, secured, and ready to use.  

The Surface Management Portal within the Intune Admin Center provides IT with a clear view of device health, compliance, and usage across all Surface devices, with Security Co-Pilot now included. IT can use AI tools to spot issues, assess risks, and respond more confidently. An actor helps safeguard on-screen information and can reduce the risk of screen exposure when working in public spaces.  

Surface brings together Microsoft’s hardware, software, and cloud to give you a solution that’s easy to manage, secure to deploy, and ready for AI from the start.  

A Linked Future, Built for Business 

Surface protects your data and privacy while helping your employees get more done.  

Surface Laptop 5G is now part of our expanding Co-Pilot + PC lineup, built for the changing needs of today’s workforce. With the 12-inch Surface Pro for business and 13-inch Surface Laptop for business now available and Surface Laptop 5G coming on August 26, organizations will have more options than ever to operate with powerful, secure, and AI-ready devices.  

When paired with Verizon’s secure, reliable 5G network, Surface devices unlock even greater productivity. The combination of Surface Co-Pilot, plus PCs, Microsoft 365 Co-Pilot, and high-speed mobile connectivity enables an effortless experience that helps businesses to work more easily and efficiently in the office or out in the field. In the U.S., Surface for Business devices are available through a broad network of partners, including Verizon Business, with selected Verizon stores rolling out in the coming months.  

Support for Windows 10 PCs ends on October 14, 2025. Now is the time for a modern Windows experience built for the air, secure by design, and ready for mobile work. Surface.com/business to find an authorized dealer or visit the Microsoft Store. When you shop at microsoft.com, you’ll get free shipping and an extended 60-day price protection and return menu. For a deeper technical dive, see the Surface Pro blog.

Source: Introducing Surface Laptop 5G: Seamless connectivity, built for business 

As of early 2026, Google Vio 3 (including 3.1) and OpenAI Sora 2 are leading the way in AI video generation. Both combine high-quality audio with video visuals. Vio 3 stands out for its sharp resolution and professional results, while Sora 2 is known for smooth, longer, story-focused clips.  

To summarize, VO3 is best for 4K professional cinematic and high-resolution marketing videos. Sora 2 works better for longer story-driven and consistent social content.  

Official Benchmarks: Google Vio 3 vs OpenAI Sora 2 

Feature	Google Vio 3/3.1 (2025/2026)	OpenAI Sora 2 (2025/2026)
Max Native Resolution	4K Open (2160p) at 60 fps via H.265/HEVC	2080p (native)
Max Duration	~8 to 10 seconds (native)	25+ seconds (up to 60s in testing)
Native Audio	Yes: dialog + SFX + music	Yes: synchronized dialog plus SFX
Frame rate	24-60 FPS	24 to 30 FPS
Aspect Ratios	16:9, 9:16, 1:1.	16:9, 9:16, 1:1
Strengths	4K Quality Color SCI Fast Mode	Longer clips, consistency, physics
Availability	Gemini app Vertex AI Flow	IOS app (invite) Sora.com API

Official Resolution and Visual Benchmarks 

• Google V03 offers native (4K capability): Veo 3 offers native (3840×2160) output, which is ideal for broadcast and high-end advertising.  

• OpenAI Sora 2 is designed to produce consistent photo-realistic video at 1080p (Full HD) with better character steadiness across shots.  

• Emirates V03 supports up to 60 fps (H265/AV1), permitting smoother motion than typical 21/30 fps models.  

Native Audio Benchmarks 

Models include synchronized multi-modal audio.  

• VO3 (Audio Accuracy): VO3 Sonidos, Effectos Sonoros y Sonidos De fondo. It excels at lip syncing for dialogue-heavy scenes.  

• Sora 2 (Audio physics): integrates sound that fits the scene’s physical dynamics, such as footsteps, and provides dialogue.  

Key Differences in Performance 

• Video length: Sora 2 creates clips often 20-25+ seconds long, whereas VR 3 is optimized for 8-second high-fidelity shots.  

• Creative Control and Editing: Sora 2 includes an editing suite for altering scenes, while Veo 3 via Google Flow focuses on video ingredients and high-quality generation.  

• Cameos vs Flow: Sora 2 introduces Cameos, which allow users to insert their likeness and voice into scenes. Vio 3 integrates with Google’s ecosystem: (YouTube Shorts/Gemini API).  

OpenAI just announced its latest AI video generation model, Sora 2. The model competes with Google’s updated Vio 3, which also promises realistic AI-generated videos. OpenAI describes Sora 2 as a major step, calling it the ChatGPT-5 movement. Google, in turn, calls Vio 3 its most advanced video language model.  

The competition in Generative Video Technology has advanced past silent clips and simple animations. OpenAI’s Sora 2 and Google’s Vio 3 each take a different approach, modulating how creators/developers and platforms apply Video AI. Both models combine video control features and serve both professionals and creators, but they vary in their strengths, focus, and availability.  

Core Focus and Positioning 

Sora 2 is built as a general-purpose video and audio model focusing on realism, accurate physics, and smooth storytelling. OpenAI presents it as a creative partner for making longer, storyboarded videos with features like multi-shot control and persistent scenes. One highlight is cameos, which let users add their own likeness and voice to scenes. Veo 3, on the other hand, is more platform-focused. Google features built-in audio and video creation, fast production, and easy sharing via YouTube shots and the Gemini API. Make it great for creators who need quick results.  

Audio Integration 

Models move beyond silent video, but they focus on different things.  

• Sora 2 adds synchronized dialog sound effects and full audio environments that match the action in each scene.  

• Veo 3 includes audio as a core part of its design, generating speech, music, and effects together.  

With Veo 3, fast creators can make shots with audio in a single step. Sora 2’s main audio strength is its strong correspondence with the visuals.  

Visual Detail and Control 

Visually, Sora 2 focuses on realism and believable motion, such as gymnastic moves and buoyancy effects that demonstrate its physics handling. Its multi-shot editing and control features help creators build smooth stories. Veo 3, in contrast, emphasizes cinematic quality, efficient workflows, and options for both 1080p and some 4K video. Veo 3 Fast is built for speed, while the main Veo 3 tier supports longer, higher-quality videos.  

Ecosystem and Integration 

Sora 2 is launching slowly through an invite-only iOS app in North America, with access on Sora.com and an API coming soon. OpenAI has mentioned a Sora 2 Pro tier for ChatGPT Pro users, which could serve both creative users and developers. Veo 3 is already part of Google’s ecosystem, offering developers access to the Gemini API and direct YouTube integration for creators. This makes Veo 3 easier to use widely, especially for social media production.  

Both companies focus on safety and on tracking the origin of content.  

• Sora 2 has a system card to address risks such as misuse of personal likeness and adds controls for tracking content in its app.  

• Veo 3 uses SynthID watermarking and YouTube’s detection tools to stop unauthorized AI content.  

Which Is The Right One For You 

If you want the realistic physics, multi-short storytelling, and cameo features, Sora 2 is the better option, though it is still hard to access. If you need speed, built-in audio, and easy sharing, especially on YouTube and through the Gemini API, Veo 3 is the most sensible choice right now.

Source: OpenAI Sora 2 vs Google Veo 3: It is more about realism and storytelling versus speed and audio integration 

Intel Panther Lake (Core Ultra Series 3) delivers a big jump in AI performance over Lunar Lake (Series 2), reaching up to 180 total platform TOPS vs 120 TOPS. It includes a 5th-generation NPU (50+ TOPS) with improved FP8 support and XE3 graphics, enabling faster, more local, and more energy-efficient generative AI tasks.  

Key AI and Platform Comparison 

• Total platform TOPS: Panther Lake offers up to 180, while Lunar Lake provides 120  

• NPU performance: Panther Lake’s NPU 5 delivers 50 TOPS, about 4% more than Lunar Lake’s NPU 4 (48 TOPS). It also achieves over 40% higher TOPS per unit area and better efficiency.  

• Data type support: Panther Lake NPU now supports FP8 natively, enabling larger, more efficient AI models.  

• Graphics AI GPU: Panther Lake uses the Xe3 (Celestial) architecture, which provides a significant boost to AI inference via XMX cores and delivers 77% faster gaming and graphics performance than Lunar Lake.  

• Model Size: Panther Lake can run large language models with 30-70 billion parameters on a device, while Lunar Lake is aimed at models with 7-8 billion parameters.  

• Architecture Panther Lake uses Intel’s 18A process technology, while Lunar Lake is built on TSMC nodes.  

Panther Lake’s fifth-generation MPU is built for higher performance in a smaller chip area. This means better AI-powered multitasking and more advanced private on-device AI features.  

With the arrival of Intel Core Ultra Series 3 processors codenamed Panther Lake, Intel has signaled a decisive step forward from the previous Series 2 generation codenamed Lunar Lake. This new platform isn’t just faster; it’s fundamentally more intelligent, built to support high-performance Co-Pilot plus PCs with on-device AI.  

Paired with Asus’ latest premium designs and software, the Panther Lake generation of Intel’s processors showcases how next-generation silicon can unlock lightweight laptops that are more capable, more efficient, and more adaptive for real-world use.  

Lunar Lake vs Panther Lake: Clearing up the Generations 

To avoid confusion, it is worth stating this clearly upfront:  

• Intel Core Intel Series 2 processors are known as Lunar Lake or Arrow Lake.  

• Intel Core Ultra Series 3 processors are known as the Panther Lake series.  

Lunar Lake (Series 2 processors) marked Intel’s first major push into AI-native laptop computing. Panther Lake (Series 3 processors) builds on that foundation, scaling AI performance, graphics capability, and platform intelligence to better meet the demands of modern AI workflows.  

From AI Ready to AI First Computing 

One of the most meaningful generational shifts between Lunar Lake and Panther Lake lies in AI computing throughput. While past generations introduced the idea of distributing AI workloads across the CPU, GPU, and NPU, Panther Lake dramatically expands the scale of what’s possible on-device.  

Delivering up to 180 platform TOPS, Panther Lake processors offer a substantial increase in power over Lunar Lake’s 120 TPOS. This jump enables faster local inference, smoother AI-assisted multitasking, and more responsive real-time features, all without relying on the cloud. For users, this means AI-powered photo enhancements, background effects, transcription, and generative tools feel instantaneous and private in seconds, even on a lightweight laptop.  

Smarter Cores, Better Efficiency for Mobile Work 

Panther Lake refines Intel’s Hybrid Architecture with updated Performance Cores (P-Cores), Efficient Cores (E-Cores), and low-power efficient cores (LP-E cores), all working together more intelligently than before, compared to Lunar Lake’s more limited core configurations. Intel Core Ultra Series 3 processors scale up to 16 total cores, allowing demanding applications and AI workloads to co-exist without compromising battery life.  

The result is a system that feels consistently fast throughout the day. Heavy creative tasks can surge when needed, while AI processes quietly run on low-power cores in the background. For professionals who expect desktop-class responsiveness from a thin-and-light laptop, this balance is crucial.  

NPU 5: Where AI Becomes Practical 

The center of Panther Lake’s AI leap is the new NPI/NPU5 architecture. This dedicated neural processing unit is designed specifically for sustained low-power AI workloads, something Lunar Lake introduced, but Panther Lake takes to a new level.  

Instead of pushing AI tasks directly onto the CPU or GPU, it will select which are better suited for the NPU and let the NPU handle them efficiently and continuously. This is what enables always-on AI features in Co-Pilot Plus PCs, from real-time collaboration enhancements to intelligent system optimization, without draining the battery or increasing fan noise. It’s a subtle change that dramatically improves day-to-day usability.  

Graphics Performance for Modern Creatives 

Panther Lake’s Xe 3 GPU architecture brings a clear boost in graphics performance. Lunar Lake set a high bar for integrated graphics, but Xe 3 goes even further with more Xe cores, better ray tracing, and improved AI upscaling.  

If you edit high-resolution video or use several external displays, Panther Lake offers smoother visuals and faster video processing. What’s impressive is that these improvements don’t come at the cost of the slim, portable designs Intel laptops are known for.  

Platform Intelligence Beyond Raw Performance 

Panther Lake processors update the whole laptop experience with faster connectivity, smarter memory support, and next-generation I/O. This helps your system keep up with changing workflows.  

With Intel Wi-Fi 7, Dual Bluetooth Core 6.0, and Thunderbolt 5, you get faster data transfers, lower device latency, and more flexible workstation setups. Panther Lake laptops feel more responsive and ready for the future than Lunar Lake systems.  

Asus AI PCs Powered By Panther Lake 

Asus uses Panther Lake processors to build laptops designed for the way people work and create today. The 2026 Asus Zen Duo is made for power users and creators. Its dual-screen design takes advantage of Panther Lake processors, which offer more cores, better graphics, and faster AI features. This makes multitasking, content creation, and AI tasks smoother than before.  

The Asus ZenBook S14 is all about easy mobility. This premium, lightweight Intel laptop delivers Panther Lake performance, an OLED screen, and a sleek design. It’s made for professionals who want built-in AI features, long battery life, and greater portability.  

Learn More. 

With Panther Lake processors, the 2026 Asus ZenBook Duo and ZenBook S14 offer faster workflows, smarter AI features, and top performance in compact designs. To find out more about these new laptops, click the button below.

Source: Intel® Panther Lake vs Lunar Lake: How Intel® Core™ Ultra Series 3 Processor Redefines the AI PC Era 

The Apple Vision Pro (M5) builds on the M2 model by boosting processing power, raising the refresh rate to 120 Hz, and improving efficiency. Instead of a new design, the M5 focuses on performance, offering up to 36% faster CPU and GPU speeds and better AI features. This leads to better visuals and a battery that lasts a bit longer.  

Hardware and Processing Changes: M5 vs M2 

• Chipset & Performance: The M5 chip has an 8-core CPU, while the M2 has 8 cores. This means the M5 is about 35% faster for single-core tasks and 36% faster for multi-core tasks in Geekbench 6 text tests. It also has more memory bandwidth (153 GB vs 100 GB) and a faster neural engine, which helps with AI tasks like quicker persona generation.  

• Visuals & Display: The OLED panels are unchanged, but the M5 now supports a 120 Hz refresh rate instead of 100 Hz. It can also show 10 more pixels in the center of your view, making visuals sharper. The M5 adds hardware-accelerated ray tracing and mesh shading, which improve graphics in games.  

• Battery & Power: The M5 comes with a 40W battery adapter, up from 30W on the M2. It also offers a bit more battery life, up to 3 hours of video playback, compared to 2.5 hours on the M2.  

• Design & Comfort, Cologne: Both headsets look the same, but the M5 usually includes a new, more comfortable dual neck band that helps spread the weight more evenly.  

• Software & ecosystem: Both devices use the same Vision OS, so the main user interface and app library are the same.  

Summary Table 

Feature	Vision Pro M2	Vision Pro M5
Processor	M2 (8-core CPU)	M5 (10-core CPU)
Refresh rate	Up to 100 Hz.	Up to 120 Hz.
Memory Bandwidth	100 GBs	153 GBs
Battery Life	~2.5 hours	~ 3 hours.
Power Adapter	30 W	40 W

Conclusion 

The M5 is focused on better performance and is a good choice for power users or developers; however, the M2 is still a solid option for most people. Some users say that changes like clearer text and faster app loading are noticeable but not dramatic.  

Apple is fully embracing its new M5 chip, adding it to all its main products, as well as updated MacBooks and iPads. Apple also revealed a new version of the Vision Pro. This is the first update to the Apple Mixed Reality Headset, which it calls a Spatial Computer.  

Since the first M2-powered Vision Pro launched last year, the main change has been the switch from the M2 chip to the more powerful M5. The headset’s design and features are mostly the same; however, Apple has made a few other improvements, so we will break down what has changed and what hasn’t.  

Processor: A Faster Chip 

This is the most significant upgrade. The new Vision Pro moves from the first M2 chip (found in the 2022 MacBook Air and iPad Pro) to the M5 chip (now in the latest MacBook Pro and iPad Pro). We haven’t tested the M5 yet, but we can now look at how performance improved between the M2 and M4 chips in the 2024 iPad Pro and 2025 MacBook Air.  

On Geekbench 6, the iPad Pro with the M2 chip scores 2,545 for single-core and 9,505 for multi-core. The M4 iPad Pro scored 3,679 for single-core and 14,647 for multi-core. That’s a big jump, and the M5 should be even faster. Of course, mixed reality headsets work differently than tablets and laptops, so it’s hard to know exactly how much faster the new Vision Pro will be. Still, the rawpower should be a noticeable upgrade.  

Both the M2 and M5 are paired with Apple’s R1 chip for spatial processing, so the actual motion-tracking and environment-scanning functions will likely remain unchanged. Individual apps should benefit from the faster CPU, and this may offer a better experience when using multiple steps, multiple apps simultaneously (especially those that primarily use 2D windows rather than spatial computing). For reference, Apple says third-party apps will run twice as fast on the new Vision Pro compared to the original model.  

Both headsets appear to be physically identical, at least in terms of the main hardware. They have smooth aluminum bodies and curved glass fronts, similar to those of the Apple Watch. Vision Pro came with a dual-loop band that supports the device across the top and back of the head, using narrow, stiff straps. A solo-knit band features a nice-of-healing wider fabric strap that runs only along the back of the head. Neither is ideal, and both make the headset feel very front-heavy. The dual-knit band is the best of both worlds, potentially better distributing the headset’s load and resolving one of the Vision Pro’s biggest design complaints.  

Display: A Faster Refresh Rate 

Won’t get a sharper or more colorful display on the new Vision Pro, as it uses the same 23 million-pixel 92% DCI-P3 macro OLED display as the original. You might move more smoothly, though, as Apple says the upgraded headset supports up to 120 Hz, up from the previous version’s 100 Hz. This seems entirely due to the M5’s power, not the display itself. More fluid and consistent movement can reduce the risk of motion sickness, which is beneficial.  

Battery: A Half Hour More 

The new Vision Pro can last even longer than the original, providing up to 2.5 hours of general use or 3 hours of video playback. It’s still not great, but that’s a good 30 minutes more than the first Vision Pro.  

What Remains the Same 

Both devices run Vision OS 26 and feature the same easy-to-use eye and hand-tracking controls. They offer the most advanced mixed reality interface so far. Storage choices are still 256 GB, 512 GB, or 1 TB. Each version comes with 18 GB RAM. It comes with 4x F/2.0 stereophonic 3D main cameras, 6 world-facing tracking cameras, and 4 eye-tracking cameras. Connectivity hasn’t changed either, with Wi-Fi 6 and Bluetooth 5.2. The price is still a steep US$3,499.  

I haven’t tried the new Vision Pro yet, but if you already own one, the upgrade will not be worth spending another $3,500 just a year after buying the first. Still, Vision Pro is the top mixed reality headset right now, and the 2025 model’s better chip is a nice improvement. I am excited to test it soon, so check back for my full review and clear buying advice.

Source: Apple M5 Vision Pro vs. M2 Vision Pro: Same Vision, More Power 

The NVIDIA Blackwell B200 is a big step forward from the Hopper H200. It delivers much bigger AI throughput and data bandwidth and is built for trillion-parameter models and large-scale AI training. At the same time, the H200 (141 GB HBM3e) focuses on improving memory capacity and bandwidth for existing infrastructure. The B200 (192GB HBM3e) features a dual-drive chiplet design that doubles performance in most metrics.  

Key Specifications Comparison 

Feature	Hopper B200	Blackwell B200
Architecture.	Hopper (monolithic)	Blackwell (dual die)
Transistor Count	~ 80 Billion	~ 208 million
Memory Capacity	141 GB, HBM 3E	192 GB, HBM 3E
Memory bandwidth	~ 4.8 TB’s	~ 8 TBs
Interconnect	NVLink4 (~0.9 TB’s)	NVLink5 (1.8 TB’s).
Peak AI (FP8/FP4)	~4PF (FP8)	~9 PF (FP 8) /18PF (FP 4).
TDP (Power)	~700 W	~1000 to 1200 W+

Data Center Throughput 

• Training performance: The B200 can deliver up to 3 times better training performance for large models than the previousgeneration.  

• Inference throughput: for AI inference, the B200 provides up to 2.5-3 times the throughput of the H200.  

• FP4 capabilities: One key advantage of the B200 is its native support for 4-bit floating-point (FP4) format. This enables up to 18-20 petaflops of performance, about twice the throughput of the H200s FP8 in some cases.  

• Transformer engine: The B200 features a second-generation transformer engine, enabling faster, more efficient AI model processing.  

Memory Bandwidth and Capacity 

• Capacity: The B200 has 192 GB of HBM3E memory, while the H200 has 141 GB.  

• Bandwidth: The B200 offers 8 TB/S of bandwidth, much higher than the H200’s 4.8 TB/S. This helps reduce memory bottlenecks during large model training.  

Interconnect and Scaling 

• NVLink 5: The B200 uses the fifth-generation NVLink, providing 1.8 TB/s of Intel GPU communication bandwidth. This is twice as fast as the H200’s 900 TB/s.  

• Scaling Cologne: The Blackwell architecture can support much larger and denser clusters, such as the GB200 and NVL72. It allows up to 576 GPUs to work together as one unit.  

Summary of Use Cases 

• Hopper H200: High-throughput inference for models under 100B parameters, upgrading existing H100 infrastructure, and cases where power cooling limitations exist.  

• Choose the Blackwell B200 for training next-generation base models running a trillion-parameter model inference, handling long-context LLMs, and using new liquid-cooled high-density data centers.  

NVIDIA’s latest GPUs can make choosing the right hardware for AI projects tricky. The H100 is a proven, reliable option. The H200 offers much more memory, and the new B200 promises huge performance gains. Still, with high prices and unpredictable availability, it is important to look past the marketing and understand what really sets these chips apart.  

We have looked at how each GPU performs in real-world scenarios, including power consumption and actual performance, to help you decide which one fits your needs and schedule.  

H200: the Memory Monster 

The H200 builds on the H100 by offering much more memory. Using the NVIDIA HAPA architecture, the H200 is the first GPU to provide 141 GB of HBM3e memory with a bandwidth of 4.8 TB/s.  

Key Specifications:  

• Memory: 141 GB HBM3e.  

• Memory Bandwidth: 4.8 TB/s  

• TDP: 700 W (same as H100)  

• Architecture: Hopper  

• Best for: larger models (100+ B+ parameters), long-context applications.  

A key advantage is that both the H100 and H200 use the same 700 W of power. The H200 is not only faster, but it also delivers higher throughput without increasing power consumption.  

B200: The Future Unleashed 

B200 is NVIDIA’s new flagship based on the Blackwell Architecture. It has 208 billion transistors compared to 80 billion in the H100 and H200, and brings major new features.  

Key specifications  

• Memory: 192 GB/HBM3E  

• Memory bandwidth: 8 TBs  

• TDP: 1000W  

• Architecture: black file (dual chip design)  

• Best for: next-gen models, extremely long contexts, future-proofing  

Performance Deep Dive: Where Rubber Meets the Road 

Training Performance: Performance data shows that a single Blackwell 200 GPU is about 2.5 times faster than a single H200 GPU, measured in tokens per second. The DJX B200 system offers three times the training performance and 15 times the inference performance compared to the DJX H100 system.  

Inference Capabilities: For organizations that value deployment, inference performance is often more important than training speed. The H200 can double the inference speed of the H100 when running large language models like Llama2. The B200 offers even greater gains, with up to 15 times the inference performance of H100 systems.  

Memory Bandwidth: The Unrecognized Hero  

Memory bandwidth affects how quickly a GPU can supply data to its compute cores. Higher bandwidth means data moves faster, which can improve performance.  

• HB100:3.35 TB/s(respectable)  

• H200: 4.8 TB/s (43% Improvement)  

• B200: 8TB/s (another universe)  

The H200’s memory bandwidth is 4.8 TB/s, an increase from the H100’s 3.35 TB/s. This extra bandwidth is important for processing large datasets, as it helps reduce wait times for data during memory-intensive tasks. This may result in faster training.  

Cost Analysis: What You Are Paying 

Pricing on these GPUs has been all over the map this year. The H100 started 2025 at around $8 per hour on cloud platforms, but increased supply has pushed that down to as low as $1.90 per hour. Following the recent AWS price cuts of up to 44%, with typical ranges of $2 to $3.5 per hour, depending on the provider  

If you plan to buy hardware directly, expect to pay at least $25,000 for each H100 GPU. After adding costs for networking, cooling, and other infrastructure, a full multi-GPU setup can easily exceed $400,000. These are major investments.  

H200 Premium, you can expect to pay about 20-25% more for the H100, whether you buy or rent in the cloud. For some workloads, the extra memory makes the higher price worthwhile.  

B200 investment: The B200 will cost at least 25% more than the H100. The H200 at first may be hard to get in early 2025; however, it offers excellent long-term performance and effectiveness. If you want the latest technology right away, you will pay a premium. 

Source: H100 vs. H200 vs. B200: Choosing the Right NVIDIA GPUs for Your AI Workload 

On Wednesday, Anthropic announced it has acquired Vercept, an AI startup closely connected to some of Seattle’s top tech figures. It follows Anthropic’s December purchase of Bunn, a coding agent engine, to help grow Claude Code.  

Vercept had created tools for more complex agentic tasks, including its product Vy, a cloud-based computer-use agent that could operate a remote Apple MacBook. Vercept is one of many startups working to reimagine the personal computer for the age of AI agents. As part of the deal, Anthropic is shuttering Vercept’s product on March 25th.  

The startup graduated from Seattle’s AI2 incubator, which the Allen Institute for AI founded. Vercept co-founders also previously worked as researchers at the Allen Union Institute. One co-founder, Matt Deitke, made headlines last year for negotiating a $250 million salary from Meta to join its super intelligence lab. On Wednesday, Deitke congratulated his former colleagues on X.  

Vercept was a well-known AI startup in the area. In a LinkedIn post about the Anthropic acquisition, CEO Kiana Ehsani said the company had raised $50 million in total. She named Seth Bannon, a board member from Fifty Years, as the lead investor. Last January, Vercept announced a $60 million seed round.  

The startup also attracted notable angel investors, including former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, and Cruise founder Kyle Vogt, and Dropbox co-founder Arash Ferdowsi, according to GeekWire.  

Anthropic’s announcement named co-founders Ehsani, Luca Weihs, and Ross Gershick as part of the team joining Anthropic; however, not all of Vercept’s co-founders are joining the company behind Claude.  

Oren Etzioni, previously named as a Vercept co-founder and investor, is well known in Seattle as the founding leader of the Allen Institute for AI, like Deitke. He is not joining Anthropic and was openly unhappy about the acqui-hire, which he posted on LinkedIn after about a year. Vercept is throwing in the towel and giving its customers 30 days to get off the platform. Sad to see a fantastic team join Anthropic. I wish them the very best.  

In his LinkedIn post, Etzioni accused Bannon-Vercept’s lead investor of being partially responsible for not hiring the right business people. The two investors argued back and forth, with Bannon replying, “You disparaged the history, the heroic work, of the founders for achieving an outcome most would only have dreamed of.” They also accused each other of things like lying and making legal threats.  

Public arguments between investors may be entertaining, but they don’t mean much. What matters is the high stakes in building the next big AI company. Now, a bright startup with significant funding will become part of Anthropic.  

The terms of the deal were not disclosed, but Etzioni says he made a return on his investment. Anthropic clearly wanted these researchers, especially since one of them is now at Meta.  

Still, Etzioni told the choir he is disappointed. I am pleased to have gotten a positive return, but obviously disappointed that, after just a little over a year, with so much traction and such a fantastic team, we are basically throwing in the towel, he said.  

We could build independently and work toward the same vision as two separate versions, or join forces with an incredible team and accelerate that vision into reality. The decision to join Anthropic was easy, she said. 

Source: Anthropic acquires computer-use AI startup Vercept after Meta poached one of its founders 

Apple is expanding its factory operations in Houston, which means Mac Mini will be made in the US for the first time. The company will also increase production of AI servers and open a new training center to help people develop new skills.  

Apple announced today that it will expand its factory operations in Houston, bringing Mac mini production to the US for the first time. The company will also boost AI Server manufacturing at the site and start hands-on training at a new Advanced Manufacturing Center later this year. These efforts are expected to create thousands of jobs in Houston.  

Apple is strongly committed to the future of American manufacturing, and we are proud to expand our footprint in Houston significantly. Said Tim Cook, Apple’s CEO. We began shipping advanced AI servers from Houston ahead of schedule, and we are excited to accelerate that work even further.  

For over 20 years, people worldwide have used the Mac Mini for its powerful performance in a small design. Its cutting-edge AI features make it an important tool for students, creators, and small business owners. Starting later this year, Mac Mini will be made at a new factory in Houston, which will double the size of Apple’s campus there.  

Apple started making advanced AI servers in Houston in 2025, and production is already ahead of schedule. The servers built in Houston, including on-site-manufactured logic boards, are used in Apple data centers across the country.  

People are also investing in people, in the people who will determine the future of American manufacturing. Later this year, the company will open a 20,000 sq ft advanced manufacturing center in Houston. This new facility, now being built, will offer hands-on training in advanced manufacturing to students, supplier employees, and businesses of all sizes. Apple experts will teach the same processes used to make Apple products, helping American manufacturers improve their skills.  

Since Apple announced its $600 million pledge to the US last year, the company and its American Manufacturing Program partners have already achieved several milestones:  

• Apple surpassed its goal by sourcing over 20 billion U.S.-made chips from 24 factories in 12 states, working with partners such as TSMC, Broadcom, and Texas Instruments.  

• Global Wafers has started making bare silicon wafers at its new $4 billion facility in Sherman, Texas. Following Apple’s guidance, these wafers will be used by Apple’s chip manufacturing partners in the US, including TSMC and Texas Instruments.  

• With Apple’s investment, Amkor has begun construction of a new $7 billion advanced semiconductor packaging and testing facility in Peoria, Arizona. Apple will be the first and largest customer at this site.  

• Corning’s Harrodsburg, Kentucky, facility now makes only Cover Glass for iPhones and Apple Watches shipped worldwide. By the end of this year, every new iPhone and Apple Watch will have Cover Glass made in Kentucky.  

• In 2026, Apple expects to buy well over 100 million advanced chips made by TSMC at its Arizona facility. This is a big increase compared to 2025.  

• Apple iPhone, with the manufacturing academy in Detroit, which already helps over 130 small and medium-sized American businesses with hands-on training in AI automation and smart manufacturing. The academy recently added new online programs so businesses across the country can access the curriculum created by Apple experts and Michigan State University faculty whenever they need it.  

The plans to move some Mac Mini production from Asia to the US later this year. Production will start at Foxconn’s Houston facility to eventually meet local demand for the Mac Mini.  

Apple’s Chief Operating Officer, Sabih Khan, announced the news in a video interview with The Wall Street Journal. We are very excited to tell you that later this year, we will begin Mac Mini manufacturing right here in this space, Khan said. Over time, we want to scale Mac Mini here to serve our customers in this area.  

According to the Journal, this step is part of Apple’s 600 billion plan to increase manufacturing in the U.S. Right now, Apple produces thousands of Mac Minis in Asia each week, but only a small number will be made in the U.S. at first.  

This is not the first time Mac computers have been made in the US. Apple started making Mac Pros at its Austin facility in 2013 and still does, but only in small numbers. Khan said production there has stayed limited because demand for the Mac Pro is low.  

The video also shows Apple chip manufacturing at Global Wafers Facility in Sherman, Texas, and at TSMC and ASML sites in Phoenix. Apple says it does not build these factories but encourages its suppliers to use them.

Source: Apple today announced a significant expansion of factory operations in Houston

 Apple Will Move Some of Its Mac Mini Production to the US Later This Year