z.ai’s Open Source GLM-5 Sets New Low in Hallucination with RL ‘Slime’ Technique
Chinese AI startup Zhupai, also known as z.ai, has unveiled its latest large language model, GLM-5. This model represents a significant step forward in the company’s ongoing GLM series, maintaining an open-source MIT License that makes it suitable for enterprise use.
GLM-5’s Impressive Achievements
One of GLM-5’s major accomplishments is its remarkably low hallucination rate. On the Artificial Analysis Intelligence Index version 4.0, the model received a score of -1, amounting to a 35-point enhancement from its predecessor, GLM-4.5. This milestone positions GLM-5 ahead of notable competitors in the AI industry, including Google and OpenAI, by demonstrating knowledge reliability and the ability to refrain from fabricating information when unsure.
High-Utility Features
GLM-5 is designed for efficiency in knowledge work. It features a native “Agent Mode” that transforms raw prompts and source materials into professional documents, such as .docx, .pdf, and .xlsx files. Its capabilities extend to generating diverse outputs, from detailed financial reports to sponsorship proposals. This functionality is integrated into enterprise workflows, making it a valuable asset for businesses.
Cost-Effectiveness
In terms of pricing, GLM-5 is disruptively affordable. It costs approximately $0.80 per million input tokens and $2.56 per million output tokens, making it up to six times cheaper than proprietary alternatives like Claude Opus 4.6.
Technological Advancements
GLM-5 boasts a substantial increase in parameters, expanding from 355 billion in GLM-4.5 to an impressive 744 billion in its latest iteration. This enhancement is supported by increased pre-training data, now totaling 28.5 trillion tokens. A novel asynchronous reinforcement learning infrastructure called “Slime” addresses training inefficiencies and optimizes the generation process. This approach allows the model to produce complex behaviors more efficiently by breaking traditional bottlenecks.
Agentic Engineering
The design of GLM-5 emphasizes what Zai refers to as “Agentic Engineering.” This framework allows users to define quality gates while the AI handles execution. The model’s efficiency in transforming high-level goals into actionable tasks is expected to redefine the standard for AI applications in professional settings.
Benchmarks and Market Position
GLM-5 has achieved remarkable benchmarks, outperforming rivals such as Gemini 3 Pro and surpassing the Chinese competitor Moonshot’s Kimi K2.5. Its performance on the SWE-bench Verified achieved a score of 77.8, approaching Claude Opus 4.6 and establishing GLM-5 as the foremost open-source AI model globally.
Cost Comparison with Competitors
Below is a summary of GLM-5’s pricing compared to other leading models:
| Model | Input Cost (per 1M tokens) | Output Cost (per 1M tokens) | Total Cost |
|---|---|---|---|
| GLM-5 | $1.00 | $3.20 | $4.20 |
| Claude Opus 4.6 | $5.00 | $25.00 | $30.00 |
| Gemini 3 Pro | $2.00 | $12.00 | $14.00 |
Considerations for Adoption
While GLM-5 presents a compelling case for adoption, organizations must consider the necessary infrastructure. With 744 billion parameters, it requires substantial hardware capabilities that may pose challenges for smaller firms. Moreover, enterprises must also contemplate the geopolitical implications of deploying a solution developed in China, particularly in regulated industries.
Users have noted that the model, despite its high performance, may lack situational awareness. This observation raises concerns about its capacity to reason and operate autonomously in complex environments.
In summary, GLM-5 serves as a formidable option for enterprises looking to enhance productivity with cost-effective, advanced AI solutions. Its open-source nature, coupled with its innovative features, provides a strategic advantage in the evolving landscape of artificial intelligence. As businesses shift towards more autonomous AI systems, GLM-5 represents a crucial tool for future growth.
