Z.ai's GLM-5.3 post-training boost closes coding gap, trails GPT-5.6 Sol
Same 743B MoE base, massive RL gains: Terminal-Bench jumps from 4.6 to 28.3.
Z.ai has launched GLM-5.3, a 743-billion-parameter Mixture-of-Experts language model built as an open-weight challenger to closed frontier systems like GPT-5.6 Sol and Claude Fable 5. The defining technical twist: the model uses the exact same pretrained base as GLM-5.2 — no new pretraining, no added parameters. All performance gains come from expanded post-training, including more training environments, broader task variety, and additional reinforcement-learning compute. This is a notable proof point that substantial capability improvements can still be extracted from a fixed base model through post-training alone.
Z.ai's own vendor-run benchmarks show dramatic progress on terminal and agentic tasks. Terminal-Bench 3.0 scores jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and ExploitGym (2h/6h) from 29/39 to 105/130. However, the model still trails GPT-5.6 Sol on Terminal-Bench (34.6), DeepSWE (72.7), and ExploitBench (where an unnamed competitor leads at 78.0% vs GLM-5.3's 54.4%). GLM-5.3 also holds a 1 million-token context window via the glm-5.3[1m] model ID, supports three reasoning-effort levels (low, high, max), and is positioned on cost and deployment control rather than raw capability ceiling. MIT-licensed open weights are expected around August 28, 2026.
- GLM-5.3 uses the same 743B-parameter MoE base as GLM-5.2, with all gains from post-training and RL compute.
- Terminal-Bench 3.0 jumps from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, and CyberGym reaches 84.5% vs GPT-5.6 Sol's 83.6%.
- Open weights under MIT license arrive August 28, 2026; 1M token context available via glm-5.3[1m].
Why It Matters
Post-training-only gains challenge the scale arms race, making powerful open-weight models cheaper and more controllable.