ZAI just dropped GLM-5.3. Same 743B base they already had, but pure post-training got coding benchmarks up 50%.
745B total, 44B active. Trained on Huawei Ascend. MIT license, open-source coming.
The wild part: Terminal Bench 3.0 went from 4.6 to 28.3 — that's 6x. And they beat every closed-source model on CyberGym for security tasks. Post-training squeeze doing the heavy lifting, not just throwing more params at the problem.
