Z.ai Launches GLM-5.3-Flash, an MIT-Licensed Coding Model That Was Secretly Topping Leaderboards as 'Ox-Alpha'

Overview

Z.ai has released GLM-5.3-Flash, a mixture-of-experts coding and agentic model, under an MIT license, according to TestingCatalog. The launch resolves a days-long mystery in developer circles: GLM-5.3-Flash is the model that had been quietly topping usage charts on OpenRouter and OpenCode under the anonymous codename “ox-alpha,” a connection confirmed both by TestingCatalog and by OpenRouter’s own model listing, which states plainly that “the stealth model ‘Ox Alpha’ was revealed to be GLM-5.3 Flash.”

What We Know

GLM-5.3-Flash is described by Z.ai as the first natively multimodal model in the GLM-5 family, according to its model card on Hugging Face. It is a mixture-of-experts model with 320 billion total parameters and 18 billion active parameters, according to both TestingCatalog and the Hugging Face model card, and it was trained on a 30-trillion-token multimodal corpus, per the same two sources. The architecture combines linear attention for local dependencies with sparse attention for global context, and the model supports a context window reaching roughly one million tokens — OpenRouter lists it precisely at 1,048,576 tokens with up to 131,072 completion tokens.

On benchmarks, the Hugging Face model card reports a score of 84.3 on Terminal-Bench 2.1 and 55.3 on Humanity’s Last Exam with tools, alongside a DeepSWE score of 63.4. TestingCatalog reports that same 63.4 DeepSWE score as an improvement over predecessor GLM-5.2’s 46.2, and a jump to 48.8 on AutomationBench from GLM-5.2’s 26.2. On the Artificial Analysis Intelligence Index, the model scored 57, ranking first among 173 models tracked by Artificial Analysis, which also lists list pricing of $0.15 per million input tokens and $0.50 per million output tokens with an 83 percent cache discount. Z.ai says the model outperforms GLM-5.2 on reported coding and agentic tests at one-tenth the price, while approaching Claude Opus 4.8 on its internal coding benchmark, according to TestingCatalog.

The ox-alpha mystery began when the model appeared anonymously on OpenRouter and OpenCode on August 20, according to OfficeChai, which reports the anonymous model ended DeepSeek’s 56-day streak atop the OpenCode leaderboard. Z.ai confirmed to Bloomberg that the model circulating as “Ox Alpha” was a new entry in its GLM series, according to Trending Topics, which reported that the anonymous model scored approximately 63 percent on the full 113-task DeepSWE run — a figure that matches GLM-5.3-Flash’s confirmed 63.4 DeepSWE score and lines up with the model’s roughly one-million-token context window and multimodal, tool-calling capabilities also cited in that report.

GLM-5.3-Flash is now available to all GLM Coding Plan subscribers with three times the usable quota of GLM-5.3, and its multimodal features are offered in Z.ai’s ZCode product through Browser Use and Computer Use modes, according to TestingCatalog. The weights are published on Hugging Face, with local deployment support for SGLang, vLLM and TokenSpeed, per the same report, and OpenRouter lists Z.ai as the direct host with support for text, image and video inputs plus tool calling and JSON-formatted responses.

What We Don’t Know

Z.ai’s comparisons to GLM-5.2 and to Claude Opus 4.8 are self-reported figures relayed by TestingCatalog; independent, third-party reproductions of these specific benchmark numbers were not available at publication time. Some coverage of the broader “Ox Alpha” episode describes pricing and DeepSWE scores that do not match GLM-5.3-Flash’s confirmed figures, suggesting other outlets may be describing a separate, non-Flash tier of GLM-5.3 rather than the Flash variant covered here; this article restricts its figures to those directly tied to GLM-5.3-Flash by TestingCatalog, Hugging Face and OpenRouter.

Context

GLM-5.3-Flash follows Zhipu’s GLM-5.2, an open-weight coding model the company released in June. Z.ai has continued releasing open-weight, MIT-licensed models in the GLM series even as US labs push proprietary coding models, positioning the anonymous top-of-leaderboard debut as a way to let the model’s benchmark performance speak before its origin was known.