GLM-5.3 Flash

On August 26, 2026, Z.ai (the international brand of Beijing lab Zhipu AI) officially confirmed that the anonymous "Ox Alpha" model which had quietly topped OpenRouter usage for six days was GLM-5.3-Flash, a natively multimodal entry in the GLM-5 family. Per the article, GLM-5.3-Flash is a 320-billion-parameter mixture The article frames GLM-5.3-Flash as strategically distinct from a typical "Flash" trim: it is built on a newly trained base rather than distilled from the flagship, arrives while Z.ai's promised GLM-5.3 744B open weights have not yet shipped, and is positioned by Artificial Analysis as the cheapest model at its Intelli

Z.ai released GLM-5.3 Flash on August 26, 2026, as an open-weight 320B-total / 18B-active MoE and the first natively multimodal model in the GLM-5 series, succeeding GLM-5.2 on a newly trained base. According to the article, it was previously previewed anonymously on OpenRouter as "Ox Alpha," ships with a 1M-token cont The piece highlights that GLM-5.3's promised 744B open weights remain unpublished, with no GLM-5.3 repository on Z.ai's Hugging Face organization, making GLM-5.3 Flash the only large open-weight GLM-5 model developers can access today. It also notes that API pricing for the Flash variant is roughly one-ninth that of GL

This head-to-head comparison reports independent Artificial Analysis testing in which GLM-5.3 scores 60 on the Intelligence Index versus 57 for GLM-5.3 Flash, yet GLM-5.3 Flash costs roughly one-ninth as much at normal API pricing ($0.15/$0.50 per million tokens vs. $1.40/$4.40) and activates only 18B parameters versus Per the spec table, both models ship from developer Z.ai with a 1M-token context and support reasoning and tool use, but GLM-5.3 Flash adds native image input and native multimodality, and its weights are publicly available now while GLM-5.3's open weights remain forthcoming. The piece concludes that GLM-5.3 is the bet

DataCamp reports that Z.ai’s GLM-5.3-Flash uses a 320B-A18B mixture-of-experts design, with 320 billion aggregate parameters and 18 billion active per token, and supports a one-million-token context window. The article describes the exact GLM-5.3-Flash variant as natively multimodal and cost-optimized relative to the l The article gives a Terminal-Bench 2.1 score of 84.3 for GLM-5.3-Flash, compared with 85.0 for Claude Opus 4.8 and 87.4 for GPT-5.6 Terra. It also reports a roughly $0.10 blended price per million tokens, about one-tenth of the cited GLM-5.3 blended price, while noting a 306 GiB FP8 checkpoint that makes lightweight lo

This technical deep dive details GLM-5.3-Flash as a 320B-parameter MoE with 18B active per token, natively trained in FP8 with a 1,048,576-token context window and the first natively multimodal model in the GLM-5 series. Weights are MIT-licensed on Hugging Face, the API is priced at $0.15 input / $0.50 output per milli On benchmarks, the article reports that GLM-5.3-Flash beats GLM-5.2 across six coding and agentic evaluations at roughly one-tenth of the price and approaches Claude Opus 4.8 on long-horizon agent tasks, while the larger GLM-5.3 flagship's 744B weights remain unreleased. Coverage also documents the "Ox Alpha" anonymous