GLM-5.3 Flash (SCX.ai)

HowAIWorks published a detailed technical breakdown on August 27, 2026 covering GLM-5.3-Flash's architecture: a 45-layer stack combining KDA linear-attention layers with NoPE sparse MLA layers, augmented by an IndexPool key-vector compression mechanism and Manifold-Constrained Hyper-Connections. The model routes each t The write-up explicitly documents context-window discrepancies: Z.ai's documentation and release notes state 1,048,576 tokens, the Hugging Face model card states 300,000, and OpenRouter advertises 1,310,720 with a 48,000-token completion cap. These figures describe different things — what the architecture supports vers

Artificial Analysis provides independent benchmarking for GLM-5.3-Flash, scoring it 46 on its Intelligence Index — well above the median of 22 for comparable open-weights models. Output speed measures 48.5 tokens per second, placing the model in the lower-percentile for speed. Pricing is $0.15 per 1M input tokens and $ Technical specifications confirm 320B total / 18B active parameters, a 1M-token context window (roughly 1,500 pages of 12-pt Arial), text-and-image input with text output, reasoning capability (this page shows the reasoning version), and MIT license with Hugging Face weights. The page compares GLM-5.3-Flash against 112

Zhipu AI / Z.ai released GLM-5.3-Flash on August 26, 2026, confirming that the previously anonymous "Ox Alpha" model trending on third-party coding platforms was this new release. The model is a 320-billion-total / 18-billion-active mixture-of-experts architecture and is the first natively multimodal model in the GLM-5 The OrcaRouter coverage provides useful launch-day context positioning GLM-5.3-Flash among the cheapest frontier-grade models available at release, with the model beating its predecessor GLM-5.2 across internal benchmarks. The page confirms Zhipu AI as the creator and treats the Flash variant as a distinct SKU from the