AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

Mysterious 'Ox Alpha' Revealed as Zhipu's GLM-5.3 Flash: First Native Multimodal, Runs on Domestic Chips

The mysterious 'Ox Alpha' model has been identified as Zhipu AI's newly released GLM-5.3 Flash, the first native multimodal model in its GLM-5 series, which is open-sourced and powered by domestic chips.

AI News Feed · 2026-08-27 06:38

According to Zhipu AI, GLM-5.3 Flash has 320 billion total parameters but only 18 billion active, yet outperforms the 753-billion-parameter GLM-5.2. It scores 57 points on the latest AA benchmark, tying Claude Opus 4.8, while its price is one-tenth of GLM-5.3, with a limited-time discount to one-twentieth, and one-fortieth of Claude Opus 4.8. The model also supports a 1M context length.

The model's real-world abilities were showcased in hands-on tests by the tech publication QbitAI, which had early access. In one test, GLM-5.3 Flash built a 3D interactive SpaceX Raptor engine webpage from a single prompt without further input. It also generated accurate subtitles for a multi-speaker video, created a full movie-commentary video from the classic film "The Myth," and transformed a mobile app UI design mockup into a working interactive shopping app. Zhipu AI separately demonstrated a 3D Blender scene the model constructed autonomously over 12 hours.

During its anonymous testing phase, the model soared on global platforms. OpenRouter reported it hit No. 1 on its first day and set a daily token usage record. OpenCode said it ended DeepSeek's 56-day streak atop its leaderboard. All these requests were served by domestic chips, according to Zhipu AI, a fact only revealed after the company claimed the model.

Zhipu credits the efficiency gains to a new architecture. GLM-5.3 Flash employs a hybrid of linear attention and sparse attention with a lightweight indexer, cutting attention computation by 3.01 times and KV cache by 4.44 times versus GLM-5.3. It was trained on a 30T-token multimodal corpus. Dedicated data pipelines for visual coding allow the model to see its output and refine it continuously. To support multimodal tasks and 1M context on domestic accelerators, Zhipu adopted an Encode-Prefill-Decode separated architecture and deeper optimizations, tripling end-to-end service performance compared to the initial baseline on the same hardware and achieving per-token costs on par with mainstream Nvidia GPUs.

GLM-5.3 Flash is now open-sourced globally and available through BigModel, Z.ai, and Hugging Face, with API access and a Coding Plan. Developers can deploy the model on domestic chips, and enterprises can independently host it. The release ties together a capable model, domestic computing power, and an open-source ecosystem, making frontier AI more accessible as an everyday tool.