Paime Morning Brief: Z.ai Open-Sources GLM-5.3-Flash Native Multimodal Model and More
On August 26, Z.ai released and open-sourced GLM-5.3-Flash (320B-A18B), the first native multimodal model in the GLM-5 series.
The model adopts a Mixture of Experts (MoE) architecture with a total of 320 billion parameters, activating approximately 18 billion parameters per token. It supports text and image inputs with a 1 million token context, and its model weights are open-sourced under the MIT license. In the Artificial Analysis Intelligence Index v4.
1.1 evaluation, GLM-5.3-Flash scored 57 points, tying with Claude Opus 4.
8 Max Effort. Z.ai's self-developed Z.
ai Code Bench v1.0 test shows that the model scored 29.0 at the highest reasoning setting, compared to 29.
5 for Claude Opus 4.8. In terms of architecture, GLM-5.
3-Flash combines sparse attention and linear attention, and introduces Manifold-Constrained Hyper-Connections (mHC) to reduce the computational overhead and KV Cache footprint of long-context reasoning. The model also integrates visual capabilities into Coding and Agent workflows, enabling it to process web pages, user interfaces, and code execution feedback in combination with Browser Use and Computer Use. Beyond software development, GLM-5.
3-Flash supports scenarios such as office work, financial research, and professional document processing. It can leverage external tools to complete analysis and modeling, generating files in formats like PPTX, PDF, DOCX, and XLSX. Currently, the model has been integrated into ZCode, GLM Coding Plan, and the official API, with model weights available on Hugging Face.
Regarding pricing, Z.ai stated that the regular API pricing for GLM-5.3-Flash is about one-tenth of GLM-5.
3, and about one-twentieth during a limited-time discount.
Perplexity AI Launches Portable Computer On August 25, Perplexity partnered with NVIDIA to launch Portable Computer, an on-device AI agent. The product supports running the core workflows of Perplexity Computer entirely on-device, achieving fully localized processing for operations like file analysis, local retrieval, and task scheduling. It only collaborates with cloud models on demand after obtaining user authorization. Regarding hardware and model adaptation, Portable Computer initially supports the NVIDIA DGX Spark running Linux, capable of directly running Qwen 3.8 27B and Perplexity's self-developed post-trained PPLX 27B model. All tasks completed locally do not consume Perplexity Credits. Moving forward, it will gradually introduce NVIDIA Nemotron 3.5 Lightning and its deeply optimized fine-tuned versions for fast response. Currently, Portable Computer is available to Perplexity Pro and Max users; support for Windows systems and PCs equipped with GeForce RTX / RTX PRO GPUs will be rolled out subsequently. Source
BenQ Launches Creative Pro PV50 Series Professional Monitors On August 25, BenQ released the Creative Pro PV50 series professional monitors, including the 32-inch PV3250U and 27-inch PV2750U models. Both monitors feature 4K UHD resolution and are designed for video editing and post-production color grading. In terms of color, the PV50 series is calibrated for Rec.709 and Gamma 2.4 workflows. It supports the Dark-Tone Clarity feature to improve shadow detail performance, reducing black crush and tone shifts. The series also supports full-screen color and brightness uniformity, as well as 16-bit 3D LUT hardware calibration, with calibration data stored internally in the monitor. For workflow features, the monitors offer tools such as Shadow Check, Marker, and Area Zoom, and come with a magnetic hood and a Hotkey Puck controller. Regarding interfaces, the PV50 series supports Thunderbolt 4, allowing video and data transmission via a single cable while providing up to 96W of reverse power delivery. Source
Google Launches Gemini 3.5 Transcribe Speech-to-Text Model On August 26, Google launched Gemini 3.5 Transcribe, a speech-to-text (ASR) model. The model supports real-time streaming transcription and pre-recorded audio transcription. It can automatically remove filler words like "um" and "uh," repetitions, and slips of the tongue, adjusting the transcription based on the user's subsequent self-corrections to generate more readable text. Google stated that the end-to-end transcription latency of Gemini 3.5 Transcribe is reduced by about 70% compared to its predecessor, Chirp 3. In the FLEURS multilingual benchmark, its streaming and non-streaming speech recognition Word Error Rates (WER) are 5.50% and 5.04%, respectively. Test results from Artificial Analysis show that the model's average WER in streaming and non-streaming scenarios is 4.0% and 2.6%, respectively. Additionally, the model improves transcription performance in noisy environments and the recognition of alphanumeric combinations like order numbers and postal codes. Gemini 3.5 Transcribe supports over 85 languages, can automatically identify languages, and handle multilingual switching. Users can also provide custom vocabularies to improve the recognition accuracy of industry terms, brand names, and proper nouns. For pre-recorded audio, the model supports speaker diarization and word-level timestamps. Developers can access the public preview via Google AI Studio, Google Antigravity, and the Gemini API. General users can experience these capabilities in the macOS Gemini App and the "Rambler" feature in Android Gboard in certain regions, with Chrome browser integration support coming later. Source
Google simultaneously launched Gemini Omni 1.1 Flash, a multimodal video generation and editing model aimed at video creation and editing scenarios. The model adds features such as scene extension, first and last frame interpolation, video reference, and high-resolution output.
When extending an existing video, the model can analyze up to 10 seconds of preceding content and continue generating in 10-second increments, reaching a cumulative maximum length of 40 seconds. Users can also specify the start and end frames, with the model generating a dynamic transition between them. Furthermore, the model supports inputting reference videos of up to 3 seconds to maintain consistency in characters, actions, and scenes.
The model also offers a 360p draft mode. Compared to the standard 720p mode, this mode can increase system throughput speed by up to 60%, with generation costs at about one-third of the latter. After confirming the content, users can generate videos in 1080p or up to 4K resolution.
Gemini Omni 1.1 Flash is billed based on video duration, with API prices for 360p, 720p, 1080p, and 4K videos at $0.03, $0.
10, $0.15, and $0.30 per second, respectively.
Currently, the model is available to developers via Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform, and is provided to Google AI Plus, Pro, and Ultra subscribers in Google Flow.
NVIDIA Launches NVHBM Custom High-Bandwidth Memory On August 26, NVIDIA announced the expansion of the NVLink Fusion semi-custom AI chip platform and introduced NVHBM high-bandwidth memory technology for custom XPUs. NVHBM utilizes NVIDIA's custom HBM base die, integrating the memory controller inside the HBM 3D stack to reduce the chip area, power consumption, and engineering verification costs when integrating HBM into custom AI accelerators. NVIDIA stated that compared to standard HBM4e, NVHBM can increase memory bandwidth by up to 30% and reduce HBM power consumption by up to 15%. By reducing the footprint of PHY and control logic on the XPU main chip side, this solution can free up to 25% of additional area for the compute die; under specific chip layouts, the available area of the main compute die can increase by up to 30%. The simplified interface design also reduces routing pressure on the advanced packaging interposer, leaving more space for compute units and dedicated acceleration modules. NVHBM targets workloads with high memory bandwidth requirements, such as large-scale model training and inference, and can be used to process data like model weights, KV cache, and activations. NVIDIA stated that it has completed the design and verification of the technology with multiple memory partners, and will subsequently provide NVHBM solutions through various suppliers to help NVLink Fusion customers shorten the tape-out and system verification cycles for custom ASICs or XPUs. In terms of partnerships, Amazon's chip design company Annapurna Labs will be the first to collaborate with NVIDIA to adopt NVHBM. The two parties previously confirmed that the next-generation AI chip Trainium 4 will support NVLink 6 and NVLink Fusion, but NVIDIA has not explicitly stated whether Trainium 4 will adopt NVHBM. Source
Microsoft Xbox Team Announces Physical Disc Digital Entitlement Feature On August 26, Microsoft's Xbox team officially announced the "Physical Disc Digital Entitlement" feature, which will enter testing first for Xbox Insider members on August 31. This feature allows players to link supported physical game discs for Xbox One and Xbox Series X to their personal Microsoft accounts, unlocking the corresponding digital play rights while retaining the native disc reading capability of physical discs. To use it, players simply need to log into their account on an Xbox One or Xbox Series X console with a disc drive, insert a compatible disc, and launch the game to claim the digital entitlement. Once linked, players do not need to insert the disc frequently for subsequent gameplay and can launch the game directly in digital form. If the title supports Xbox Play Anywhere (XPA) or Xbox Cloud Gaming, players can further enjoy cross-platform play on Windows PCs, handhelds, and cloud gaming platforms. Regarding compatibility and licensing mechanisms, Microsoft stated that the first batch will cover thousands of titles, with the compatibility list expanding over time. Each physical disc corresponds to only one "revocable" digital license. Source
Other Quick News Apple announced that it will hold a new product launch event at 1:00 AM Beijing Time on September 10. Source GTA 6 released a new game trailer. Source On August 27, Audio-Technica's ATH-CKD30C USB-C wired in-ear headphones officially went on sale in the US market. The ATH-CKD30C features a USB-C plug-and-play design, allowing direct connection to computers, tablets, and smartphones, with an integrated microphone in the in-line remote module. In terms of acoustic specs, the headphones are equipped with a 10mm dynamic driver and a built-in DAC, offering a frequency response of 5Hz to 25kHz and supporting up to 96kHz/24-bit USB audio input. They also preset two EQ modes: Bass Boost and Clear Voice. For wearing and configuration, the ATH-CKD30C adopts a sealed in-ear structure for passive noise isolation, but does not support ANC active noise cancellation. The headphones come with XS, S, M, and L ear tips. Currently, the US version is available in black and white, with some retail channels listing it at $224.99. Source Ricoh Imaging announced on August 27 that it is developing the new generation high-end compact digital camera RICOH GR IVx. The new camera continues the GR series design philosophy of high image quality, lightweight portability, and quick snapping, and is equipped with a newly developed 40mm equivalent GR lens, complementing the focal length of the GR IV which uses a 28mm equivalent wide-angle lens. For imaging configurations, the RICOH GR IVx features a 26.1mm ƒ/2.8-16 lens with a 35mm equivalent focal length of approximately 40mm. The official statement notes that its field of view is closer to the natural human eye, suitable for documentary and daily street photography. The body houses an approximately 25.74-million-effective-pixel APS-C CMOS sensor, supports five-axis sensor-shift image stabilization, and has a native ISO range of 100-204800. Regarding release and product line, the RICOH GR IVx is scheduled to launch in the winter of 2026 or later, with pricing to be determined. Meanwhile, due to the discontinuation of some components, the current GR IIIx, GR IIIx HDF, and GR IIIx Urban Edition are expected to gradually cease production after completing shipments in Japan in October this year. Source On August 27, according to Windows Latest citing sources familiar with Microsoft's Windows modernization plan, Microsoft still plans to gradually migrate remaining features in the Control Panel to Windows Settings or other modern interfaces. However, the Control Panel will not be directly removed until compatibility transformations for traditional components like printers, network devices, drivers, and enterprise software are completed. Reportedly, Microsoft is already testing new versions of the "Run" and "AutoPlay" dialogs rebuilt with WinUI in Windows Insider, and the new "Print Management" app has entered the Release Preview Channel. The traditional "Properties" window of File Explorer is also undergoing modernization. These components are expected to gradually replace existing Win32 or MMC legacy interfaces in subsequent Windows 11 updates. Source Google announced it will tighten performance and device-transfer experience requirements for Android apps. Android 17 will set memory limits for apps based on device RAM, which is currently enabled on Pixel devices and planned for gradual rollout to more Android manufacturers. Upon reaching the limit, the system may compress memory, limit performance, or terminate the process. Starting in February 2027, Google Play will check apps' dynamic memory, Bitmap memory, and DEX code optimization. Code optimization, shrinking, and obfuscation coverage must all reach at least 25%. Non-compliant apps and games may face reduced visibility, and some publishing features may be restricted. Starting in April 2027, apps supporting account login must also integrate Zero-Tap Sign-In. After a user changes devices and restores data, apps should automatically restore the login state via the Restore Credentials API. Games are currently exempt from this requirement. Source
Recent Updates from Sspai A new season of Sspai Premium has set sail, with updated benefits, more surprises, and a physical commemorative card. Click to learn more The Dot Skill, which allows AI assistants to directly interact with your Quote/0 e-ink screen via natural language commands, is now live. Click to learn more
Good Articles You Might Have Missed 💻 A mini laptop from 1997: Toshiba Libretto 50M nostalgic notes 🐧 The sixth year of daily Linux use: hoping these experiences ease your concerns 🌟 New Stuff 250 | What have Sspai editors bought recently? > Download the Sspai 2.0 client and follow the Sspai WeChat official account to unlock a brand new reading experience 📰 > Practical, genuine software presented by Sspai 🚀
