zai-org / GLM-5

2026-02-09 · Source: Github Trending: All languages · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

zai-org has released the GLM-5 series of large language models, including GLM-5.2, GLM-5.1, and GLM-5, designed for complex systems engineering and long-horizon agentic tasks. GLM-5.2, the latest flagship, offers a solid 1M-token context and advanced coding capabilities, achieving 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, nearing Claude Opus 4.8's 85.0. It features an improved architecture with IndexShare, reducing per-token FLOPs by 2.9×, and an MTP layer for speculative decoding. GLM-5.1 focuses on agentic engineering, demonstrating sustained effectiveness over long sessions and iterative problem-solving. The foundational GLM-5 scales to 744B parameters (40B active) from 355B (32B active) and uses 28.5T pre-training tokens, integrating DeepSeek Sparse Attention and a "slime" asynchronous RL infrastructure for efficient post-training. All models are available in BF16 and FP8 precision and support local deployment via SGLang, vLLM, Transformers, KTransformers, and Ascend NPU.

Key takeaway

For AI Engineers developing long-horizon agentic systems, you should evaluate the GLM-5 series as a robust open-source foundation. GLM-5.2 provides a solid 1M-token context and strong coding, while GLM-5.1 excels in sustained iterative problem-solving. Consider deploying these models locally with frameworks like vLLM or SGLang, and experiment with the `reasoning_effort` parameter to fine-tune performance and latency for your specific agentic workflows.

Key insights

The GLM-5 series advances large language models for long-horizon agentic tasks through architectural and training innovations.

Principles

• Scaling model parameters and pre-training data enhances AGI intelligence efficiency.
• Asynchronous RL infrastructure improves LLM training throughput and post-training iteration.
• IndexShare architecture reduces per-token FLOPs in sparse attention layers.

Method

IndexShare reuses indexers across sparse attention layers, reducing FLOPs by 2.9× at 1M context. The "slime" asynchronous RL infrastructure improves training throughput. GLM-5 models allow controlling thinking effort via `reasoning_effort` or disabling it.

In practice

• Deploy GLM-5 models locally using vLLM or SGLang for inference.
• Adjust `reasoning_effort` to optimize GLM-5 performance-latency trade-offs.

Topics

• GLM-5 Series
• Agentic LLMs
• Long Context Windows
• Sparse Attention
• Asynchronous RL
• Model Deployment
• Coding Benchmarks

Code references

• zai-org/GLM-5
• THUDM/slime
• sgl-project/sglang
• vllm-project/vllm
• huggingface/transformers

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• A Review of GLM-5: 744B-Parameter LLM Built for 200K Context and Agentic Training — GLM-5 bridges prompt-driven coding to agentic engineering via systems-level LLM advancements.
• 🗞️ Z.ai releases GLM 5.2 model: 1M context window with MIT-licensed open weights, long-horizon coding agents — Z.ai's GLM-5.2 combines a 1M-token context with novel architectural and decoding optimizations for efficient, long-horizon AI coding.
• GLM-5.2: Built for Long-Horizon Tasks — GLM-5.2 delivers reliable 1M-token context for long-horizon coding, leveraging architectural and inference optimizations.
• 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.
• 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.
• The Sequence AI of the Week #781: The Amazing GLM 4.7 — GLM-4.7 shifts open-weight LLMs from conversational ability to agentic reliability for autonomous workflows.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Github Trending: All languages.