z.ai's open source GLM-5 achieves record low hallucination rate and leverages new RL 'slime' technique

2026-02-11 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, medium

Summary

Zhupai (z.ai), a Chinese AI startup, has released GLM-5, an open-source large language model under an MIT License, featuring a record-low hallucination rate of -1 on the Artificial Analysis Intelligence Index v4.0. This represents a 35-point improvement over its predecessor, surpassing major U.S. competitors in knowledge reliability. GLM-5, with 744B parameters and a Mixture-of-Experts architecture (40B active per token), is designed for high-utility knowledge work, offering native "Agent Mode" capabilities to generate professional office documents like .docx, .pdf, and .xlsx files directly from prompts. The model is disruptively priced at approximately $0.80 per million input tokens and $2.56 per million output tokens, making it significantly more cost-effective than proprietary alternatives like Claude Opus 4.6. Its training leverages a novel asynchronous reinforcement learning infrastructure called "slime" to enhance efficiency for complex agentic behaviors.

Key takeaway

For CTOs and VPs of Engineering evaluating LLM adoption, GLM-5 offers a compelling open-source option for building autonomous office solutions. Its low hallucination rate, native document generation, and aggressive pricing make it ideal for enterprises ready to move beyond simple copilots. However, you must weigh the substantial hardware requirements, geopolitical considerations for a China-based model, and the need for robust agent-specific permissions to mitigate governance risks associated with autonomous operations.

Key insights

GLM-5 sets new benchmarks for open-source LLMs in hallucination reduction and agentic document generation at a competitive price.

Principles

• Prioritize knowing when to abstain over fabricating information.
• Agentic AI should deliver ready-to-use, formatted documents.
• Asynchronous RL can overcome long-tail training bottlenecks.

Method

The "slime" RL infrastructure uses independent trajectory generation and system-level optimizations like Active Partial Rollouts (APRIL) to accelerate training for complex agentic tasks, integrating Megatron-LM and SGLang.

In practice

• Utilize GLM-5 for direct generation of .docx, .pdf, and .xlsx files.
• Implement GLM-5 for cost-effective agentic engineering workflows.
• Consider "slime" for optimizing large-scale RL training.

Topics

• GLM-5
• Large Language Models
• Reinforcement Learning
• AI Hallucination
• Agentic AI

Code references

• THUDM/slime

Best for: CTO, VP of Engineering/Data, Entrepreneur, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Zhipu AI's GLM-5.2 closes in on closed-source leaders in coding marathons — Zhipu AI's GLM-5.2 sets a new open-source benchmark for long-horizon coding with a 1-million-token context and novel architectural efficiencies.
• OpenAI just dropped the limited preview of its new GPT 5.6 model suite — OpenAI's GPT-5.6 models, particularly Sol, demonstrate advanced agentic capabilities and cybersecurity performance but also exhibit increased risks of autonomous, boundary-crossing behavior.
• Am I Even Needed Anymore? GLM-5, Agentic Loops & AI Productivity Psychosis - EP99.34 — New AI models, particularly GLM-5, are intensifying work and shifting focus to agentic loops and non-US hardware.
• The Sequence AI of the Week #813: Deep Diving Into the Amazing GLM-5 — AI is shifting to autonomous "agentic engineering" requiring advanced reasoning and context handling.
• 🤖AI Agents Weekly: GPT-5.3-Codex-Spark, GLM-5, MiniMax M2.5, Recursive Language Models, Harness Engineering, Agentica, and More — New agentic coding and reasoning models demonstrate self-development capabilities and significant performance gains.
• Chinese AI lab Zhipu releases GLM-5 under MIT license, claims parity with top Western models — GLM-5 is an open-source, 744B-parameter Chinese AI model claiming parity with top Western models on agent and coding tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.