Z.ai introduced GLM 5.3 — a 743 billion parameter model with the same architecture size as GLM 5.2, but with scaled post-training through reinforcement learning on real multi-stage tasks. The performance increase was achieved exclusively through RL: an increase from 4.6 to 28.3 on Terminal Bench 3.0, an improvement to 66.9 on DeepSWE v1.1, as well as the unexpected emergence of cyber capabilities — the model found 2,436 vulnerabilities in 269 open-source projects.
What happened
Z.ai released GLM 5.3. The architecture and number of parameters remained unchanged — 743 billion, the same as GLM 5.2. The entire increase was obtained through scaled post-training via reinforcement learning on real long-horizon tasks: ML infrastructure optimization, working with clusters, terminal debugging. Benchmark results: Z.ai Code Bench showed a 50% increase, Terminal Bench 3.0 increased from 4.6 to 28.3, DeepSWE v1.1 — from 46.2 to 66.9. Modeling revealed emergent cyber capabilities that were not an explicit training goal: the model discovered 2,436 vulnerabilities in 269 open-source projects, including the Linux kernel and WebKit, of which 1,097 are medium and high level — the oldest vulnerability dates back to 1981. The model weights will be open-sourced approximately 2 weeks after passing a security review. The API interface has been changed: three levels of reasoning are now available — low, high, max — and thinking cannot be disabled, which represents a breaking change for legacy applications.
Context
The model update occurred without changing the architecture and without additional fine-tuning on new data — all progress was obtained exclusively through RL on real, not synthetic, tasks. The six-fold jump on Terminal Bench 3.0 indicates a qualitative transition from episodic successes to a systematic ability to interact with the terminal. Emergent cyber capabilities are a byproduct of RL on long horizons with access to real systems, which confirms the hypothesis that large-scale RL training develops generalized abilities for analysis and research. The two-week delay in opening the weights due to security is a signal that the industry recognizes the risks of uncontrolled access to models capable of planning full exploitation chains of vulnerabilities.
Why this matters for the industry
Z.ai demonstrates that scaling RL post-training on real multi-stage tasks yields breakthrough results that outperform the approach with synthetic benchmarks. If the results are reproducible, RL on long horizons will become a standard stage of post-training for coding models, and the industry will move from demonstration metrics to evaluation on production-relevant workloads. The 2,436 vulnerabilities discovered in OSS projects show that models with RL post-training can become a new class of cybersecurity tools. A window opens for startups for products at the intersection of devtools and cybersecurity. Mandatory thinking with depth levels may become an industry UX pattern.
Why this matters for users
GLM 5.3 is available through the Z.ai API right now — you can start testing coding agents and cybersecurity tools. The breaking change with mandatory thinking will require refactoring integrations: all clients will experience increased latency and request costs. Without public data on pricing and latency, it is impossible to calculate the final cost of migration, so it is recommended to test in a staging environment with real metric measurements before deploying to production. The weights will be open-sourced in ~2 weeks — after that, the possibility of self-hosting the model and the emergence of an open-source tool ecosystem will appear. The model is capable of independently performing multi-day tasks of ML cluster optimization and terminal debugging.
What is still unknown / limitations
There is a lack of public data on the reward functions and sample budget used for Z.ai's RL post-training, which makes attempts to reproduce the methodology premature. There is no information on API pricing and latency — early integration into production without these metrics is speculative. There is no breakdown of DeepSWE v1.1 results by task categories, which makes the picture incomplete. The stated API availability "right now" does not guarantee production-readiness — independent validation on real workloads is required.
Sources
- GLM-5.3: Frontier Coding with Emergent Cyber Capabilities — official Z.ai blog
- Z.ai Launches GLM 5.3 — Unite AI
- GLM-5.3 — Benchmarks, Specs & Release Date — AI Release Tracker
Author
Look at AI, editorial team