Z.ai releases GLM-5.3-Flash and confirms its Ox Alpha identity
Z.ai releases GLM-5.3-Flash with open weights and confirms it was Ox Alpha. It leads DeepSeek-V4-Vision-Exp on six of eight listed coding and agent benchmarks.
What happened
Event details
Z.ai released GLM-5.3-Flash on August 26, 2026, confirming that it had been tested anonymously as ox-alpha on OpenCode and OpenRouter before release. The first natively multimodal model in the GLM-5 family has 320B total parameters, 18B active parameters and a one-million-token context window. It combines sparse and linear attention, with weights released under the MIT license. In Z.ai's comparison table of eight coding and agent benchmarks, GLM-5.3-Flash leads DeepSeek-V4-Vision-Exp on six and trails on two. Scores include 84.3 versus 83.9 on Terminal Bench 2.1, 63.4 versus 59.3 on DeepSWE v1.1 and 78.4 versus 75.9 on Toolathlon Verified. It trails at 56.3 versus 57.7 on NL2Repo and 26.3 versus 27.3 on Agents' Last Exam. The results support a slight overall advantage rather than a lead on every benchmark.
Assessment
Why it matters
The model brings native GLM-5 vision capabilities, one-million-token context and hybrid sparse/linear attention to an open-weight Flash variant for the first time. Its anonymous Ox Alpha test attracted substantial developer adoption, while long-term impact depends on deployment support and sustained use.
Availability
Access notes
Available through the Z.ai API and GLM Coding Plan, with public model weights on Hugging Face.