Chinese AI company Zhipu AI officially unveiled its new open-weights model GLM-5.3-Flash on August 26, 2026, releasing the weights under a permissive MIT license. The launch follows weeks of speculation across the global developer community regarding a high-performance, mystery model that had appeared on benchmarking platforms under the pseudonym Ox Alpha. With this release, Zhipu AI makes a system specifically optimized for high throughput and cost-effective deployment freely accessible to developers and enterprises.
Architecturally, GLM-5.3-Flash employs a Mixture-of-Experts (MoE) design with 320 billion total parameters. However, only 18 billion parameters are activated per individual inference step, significantly curbing the computational footprint per generated token. At the same time, the model supports a long context window of up to one million tokens. To manage the immense key-value cache overhead typically associated with large context lengths, Zhipu AI integrated a hybrid approach combining sparse and linear attention mechanisms.
Prior to its formal unmasking, the model generated substantial momentum within the AI scene. Running on inference platforms such as OpenRouter and OpenCode under the codename Ox Alpha, the system processed more than 62 trillion tokens in anonymous production settings. Many developers praised its rapid latency and solid reasoning capabilities without knowing the underlying vendor or architectural specifications.
A key element of the release centers on the computing infrastructure used to power the deployment. According to Zhipu AI, the entire inference traffic during the testing period was handled by a massive cluster comprising 100,000 domestically manufactured Chinese AI accelerator chips. The reliable operation of this cluster highlights ongoing advances in China's hardware ecosystem to support production-scale inference independently of Western supply chains.
By opting for an MIT license, Zhipu AI grants users wide latitude for commercial deployment, fine-tuning, and on-premises integration. The blend of reduced serving costs, parameter efficiency, and full open-weights availability is poised to intensify competition among open foundation models designed for long-context workloads.