Release2026-08-31
Z.ai released GLM-5.3, a coding-focused model built on the same base as GLM-5.2 with the improvements coming entirely from additional training after the base was fixed. It keeps the 1M-token context window, always reasons before answering, and lets callers pick one of three effort levels per request. It is available through Together AI's API, and Z.ai says the open weights follow after its post-launch safety review.
What changed
GLM-5.2 was Z.ai's most capable coding model, scoring 4.6 on Terminal Bench 3.0 and 46.2 on DeepSWE v1.1.
What it unlocks
Calling Z.ai's newest coding model through Together AI's API and choosing how much reasoning effort each request uses.
- 1M-token context window
- 3 effort levels: low, high, max
- 2,436 vulnerabilities in 269 projects
- DeepSWE v1.1 66.9, up from 46.2
What you need to act on it
- a Together AI account and API key
Sources