By ByteBot
Z.ai shipped GLM-5.3 on August 14 without touching the base model. Same 743-billion-parameter mixture-of-experts architecture as GLM-5.2. Same weights, same tokenizer. The only thing that changed was post-training — and Terminal-Bench 3.0 coding scores went from 4.6% to 28.3%. That jump is worth paying attention to, because it says something bigger than “new model released.”
What Z.ai Actually Changed
The company’s framing was direct: “Scaling post-training is all we did for GLM-5.3.” That’s an unusual thing to say at a model launch. Most labs hide the kitchen, not the recipe.
The method is Reinforcement Learning with Verifiable Rewards (RLVR) — the same technique DeepSeek-R1 popularized. Z.ai’s pipeline takes it further: research agents pull task patterns from real developer workflows and convert them into runnable long-horizon environments. A judge agent then verifies each task is actually solvable before it enters the RL training mix. You get a training set of hard, real-world coding problems, confirmed to have solutions before any compute is spent on them.
The result: a model that didn’t get rebuilt but is meaningfully better at the tasks developers actually run. That’s the proof-of-concept Z.ai is shipping alongside the model.
The Benchmark Picture
On Terminal-Bench 3.0 — the current standard for evaluating long-horizon coding agents — GLM-5.3 scores 28.3%, up from 4.6% on GLM-5.2. That’s a 23-point gain from post-training alone. GPT-5.6 Sol sits at 34.6% and Claude Fable 5 at 33.7%. GLM-5.3 isn’t first, but it’s now in the same conversation.
The cybersecurity numbers are more interesting. On CyberGym — the benchmark for defensive vulnerability discovery — GLM-5.3 hits 84.5%, narrowly ahead of Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). Margins that tight are within run-to-run noise, so call it a three-way tie at the top. Where GLM-5.3 lags is offensive exploitation — ExploitBench and ExploitGym both favor Sol. Post-training made GLM-5.3 sharper at working from specifications. It didn’t turn it into a general-purpose cybersecurity powerhouse.
Z.ai also reports a 50% gain on its internal Code Bench and significant improvements on SWE-Marathon, which tests long-horizon tasks spanning hours and hundreds of tool calls. That’s the model’s actual sweet spot: multi-step, multi-hour agentic coding work — not single-shot benchmark questions.
What Developers Can Do Right Now
The API is live. Pricing held at GLM-5.2 levels: $1.40 per million input tokens, $4.40 per million output, $0.26 for cached input. Check the Z.ai developer docs for GLM-5.3 for full API reference. If you’re on the GLM Coding Plan ($18/month), you were already upgraded — no action needed.
Open weights are a different story. Z.ai expects to release them around August 28 after completing safety evaluation. Given that GLM-4.6 shipped under MIT license, expect similar terms. A 743B MoE model with a 1-million-token context window and roughly 40B active parameters per token is expensive to run, but the IndexShare optimization — reusing the attention indexer across every four sparse layers — cuts per-token compute at long context significantly.
What’s Still Missing
The June community poll had vision as the top feature request for this release. GLM-5.3 doesn’t ship it. Z.ai hasn’t explained the timeline, and reactions were predictably mixed. For a model competing with GPT-5.6’s Sol and Terra tiers — both of which handle multimodal input — the gap matters for production use cases involving screenshots, diagrams, or UI work.
The Bigger Signal
Post-training compute now exceeds pretraining spend at frontier labs. ACL 2026 documented power-law scaling behavior in RL post-training — the more you scale it, the more predictably performance improves. Z.ai’s GLM-5.3 is a public demonstration that this works at the 743B scale.
The implication for the open-model ecosystem: base weights matter less than the post-training recipe. A lab that can’t afford to pretrain from scratch might still compete at frontier coding levels if it can run the right post-training pipeline long enough. GLM-5.3 is the clearest example yet of that bet paying off.
Open weights drop around August 28. That’s when the real test starts — when the community gets to run the same RLVR-trained model and decide whether the benchmark gains hold in the workflows that matter to them.
