Z.ai has released GLM 5.3, an open weights model that achieves major performance leaps solely through post training. The new model shares the exact same base architecture as GLM 5.2. But the coding capabilities are far stronger, and the official open weights will arrive within 14 days once safety testing concludes.
Instead of building a new model from scratch, the engineering team expanded the existing software stack. They scaled up training environments to simulate actual engineering workflows rather than simple classroom exercises. The model can now access storage systems, compute clusters, and internal codebases to diagnose training bottlenecks. In internal testing on the Z.ai Code Bench, GLM 5.3 solved 34.5% of tasks at maximum effort, outperforming older editions while using fewer output tokens.
According to technical details released on the Z.ai official site, the model uses reinforcement learning strategies to maintain these gains over long horizon tasks. It can work for days on a single machine learning infrastructure problem without needing human supervision at every step. This makes it highly effective for software engineering teams who want an agent that can take ownership of a problem from start to finish.
A surprising outcome of this scale is the sudden emergence of advanced cyber security capabilities. GLM 5.3 did not just get better at finding simple software bugs. It now plans multi step attacks across entire exploitation chains. On the CyberGym evaluation, the model scored 84.5%, beating rivals like GPT 5.6 Sol. Testing on actual codebases uncovered 2,436 vulnerabilities across 269 open source projects, some of which had been hidden for over 40 years. To keep these disclosures transparent, Z.ai launched a public ledger tracking every vulnerability as it goes through the security process.
The entire system runs on slime, an open source post training framework designed for reinforcement learning. The framework aligns the training and rollout pathways mathematically. This keeps the log probability difference at an incredibly small 0.0000001 level. System optimizations also improve memory caching, allowing teachers to swap dynamically without dedicated services. The bottom line is a 2.3 times speedup in training throughput for complex coding tasks.
Developers using the API will notice some immediate changes to how the model functions. You can no longer turn off the thinking parameter. Instead, calls require selecting between low, high, or max reasoning effort. The coding plan now uses a points program where off peak usage during nights and weekends is discounted by 50%. This update is live now for all existing subscribers.
