Zhipu Open-Sources GLM-5.3-Flash; Google, NVIDIA, Perplexity Announce AI Updates

Zhipu said the model uses a Mixture-of-Experts architecture with 320 billion total parameters, activating 18 billion per token. It supports text and image input and a 1 million-token context window, and its weights are released under the MIT license. On the Artificial Analysis Intelligence Index v4.1.1, GLM-5.3-Flash scored 57, tying with Claude Opus 4.8 Max Effort. In Zhipu's own Z.ai Code Bench v1.0, it scored 29.0 at the highest reasoning level, compared with 29.5 for Claude Opus 4.8. The model combines sparse and linear attention and introduces manifold-constrained hyper-connections to reduce long-context compute and KV cache usage. It integrates vision into coding and agent workflows, enabling it to interact with web pages, user interfaces and code execution feedback via Browser Use and Computer Use. It can also generate office documents such as PPTX, PDF, DOCX and XLSX. The model is available on ZCode, GLM Coding Plan and the official API, with weights on Hugging Face. Zhipu said the API price is about one-tenth of GLM-5.3's regular rate, and one-twentieth during a limited-time discount.

Google launched two AI models on Aug. 26. Gemini 3.5 Transcribe, a speech-to-text model, supports streaming and pre-recorded audio, removes filler words and repeated content, and adjusts transcriptions based on the speaker's self-corrections. Google said end-to-end latency is about 70% lower than its predecessor Chirp 3. In FLEURS multilingual benchmarks, the model achieved word error rates of 5.50% for streaming and 5.04% for non-streaming; Artificial Analysis measured 4.0% and 2.6%, respectively. It supports more than 85 languages and custom vocabulary for domain terms. For pre-recorded audio, it offers speaker diarization and word-level timestamps. Google also introduced Gemini Omni 1.1 Flash, a multimodal video generation and editing model with scene extension, first/last frame interpolation, video reference and high-resolution output. It can extend a video in 10-second increments up to 40 seconds, based on up to 10 seconds of preceding content, and can generate transitions between given start and end frames. A 360p draft mode runs at up to 60% higher throughput and one-third the cost of the 720p mode, with final output up to 4K. API pricing ranges from $0.03 to $0.30 per second of video.

NVIDIA on Aug. 26 expanded its NVLink Fusion semi-custom AI chip platform with NVHBM, a high-bandwidth memory technology for custom XPUs. NVHBM uses a custom HBM base die and integrates memory controllers into the HBM 3D stack, reducing area, power and engineering validation costs. Compared with standard HBM4e, NVIDIA said NVHBM offers up to 30% higher bandwidth and 15% lower power consumption, and can free up to 25% of compute-die area, or about 30% with certain layouts. Annapurna Labs, Amazon's chip design unit, will be the first partner to adopt NVHBM. NVIDIA did not say whether the Trainium 4, previously confirmed to support NVLink 6 and NVLink Fusion, will use NVHBM.

Perplexity, in partnership with NVIDIA, launched Portable Computer on Aug. 25. The on-device AI agent runs Perplexity Computer's core workflows locally, including file analysis, local retrieval and task scheduling. It initially supports NVIDIA's DGX Spark with Linux, running Qwen 3.8 27B and Perplexity's own PPLX 27B models without consuming Perplexity Credits. Support for Windows and GeForce RTX/RTX PRO PCs is planned. The device is now available to Perplexity Pro and Max users.

In other product news, BenQ introduced the Creative Pro PV50 series of 4K monitors for video editing and color grading, with models in 32-inch and 27-inch sizes. Microsoft announced that Xbox Insider members will start testing a "Physical Disc Digital Entitlement" feature on Aug. 31, allowing players to link physical games to their accounts and unlock digital versions. Apple scheduled a product launch event for Sept. 10 Beijing time, and Rockstar released a new trailer for GTA 6. Audio-Technica began U.S. sales of its ATH-CKD30C USB-C earphones, with some retailers pricing them at $224.99. Ricoh Imaging said it is developing the GR IVx compact camera with a 40mm-equivalent lens, expected in winter 2026 or later.