GLM-5 from z.ai Reduces Its Hallucination Rate and Introduces RL Technique

GLM-5 from z.ai introduces significant advancements in AI model

z.ai, a Chinese artificial intelligence startup, has launched its latest large language model, GLM-5, focused on reducing failures in information generation. The open-source model, licensed under the MIT license, is particularly aimed at business use and has achieved an unprecedented hallucination rate of -1 on the Artificial Analysis Intelligence Index v4.0.

With a 35-point improvement over GLM-4.5, GLM-5 stands out for its ability to recognize situations where it should refrain from generating inaccurate information. This approach results in greater knowledge reliability, surpassing competitors like Google, OpenAI, and Anthropic.

Additionally, GLM-5 features native "Agent Mode" capabilities, allowing for direct creation of professional documents from commands or source materials, generating files in .docx, .pdf, and .xlsx formats.

Model accessible for the business market

With a cost of about $0.80 per million input tokens and $2.56 per million output tokens, GLM-5 positions itself as a low-cost option compared to proprietary models, such as Claude Opus 4.6, which costs about six times more.

Technological advancements and architecture of GLM-5

GLM-5 represents a significant advancement in its architecture, increasing from 355 billion parameters of GLM-4.5 to 744 billion parameters. The model employs a Mixture of Experts (MoE) architecture, activating 40 billion parameters per token and enabling the processing of a pre-trained data volume of 28.5 trillion tokens.

z.ai has developed the "slime" technique as a reinforcement learning (RL) infrastructure to address inefficiencies in large-scale training. This innovative approach allows the independent generation of trajectories, improving iteration for complex tasks.

The model includes optimizations such as Partial Rolling Activation (APRIL), aimed at reducing the time typically consumed in RL training.

Practical capabilities of GLM-5

Positioning itself as an office tool for the era of General Artificial Intelligence (AGI), GLM-5 is designed to generate ready-to-use documents rather than just text snippets. This functionality allows breaking down high-level goals into actionable subtasks, optimizing work for organizations seeking autonomy.

Superior performance compared to competing models

GLM-5 is considered the most powerful open-source model currently available, surpassing Chinese competitors like Kimi K2.5. The model achieved a score of 77.8 on SWE-bench Verified, surpassing Gemini 3 Pro (76.2) and closely matching Claude Opus 4.6 (80.9).

Companies across various sectors should consider adopting GLM-5, which offers a level of flexibility and access to cutting-edge intelligence without the restrictions imposed by closed-source competitors. The possibility of self-hosting intelligence may provide a decisive strategic advantage.

Implications and security considerations

However, the scale of GLM-5, with 744 billion parameters, requires robust infrastructure, which may pose a challenge for smaller companies. Additionally, concerns over the model’s origin, developed by a lab in China, should be evaluated, especially in regulated sectors.

The introduction of autonomous agents also raises governance issues, with increased risks of errors as AIs perform tasks without human supervision. Therefore, it is crucial for organizations to establish adequate quality barriers before implementing GLM-5.

Finally, GLM-5 represents not only an economical option but also a bet on the future, where the most valuable AIs will be those capable of executing tasks independently, enhancing the efficiency of organizational processes.