[glm-5.3-flash]

August 31, 2026

• OrcaRouter published refusal-removed weights for GLM-5.3-Flash (320B/18B active, native block-FP8), reporting refusal rates collapsing from 93–97% to 11–18% across MaliciousInstruct, JailbreakBench, AdvBench and HarmBench, while noting refusal does not go uniformly to zero and does not appear mediated by a single linear direction. The team defended the release after criticism over cyber-harm potential (@OrcaRouter). · AI & Model Security

in Fully Patched, Still Domain Admin

August 27, 2026

• Z.ai shipped GLM-5.3-Flash — natively multimodal with a 1M-token context — and confirmed the stealth "Ox Alpha" model is GLM-series, with weights to follow (Bloomberg) (discussion). · Frontier Models

in When the Sandbox Isn't a Boundary