Code 2025
Z.ai GLM 4.6: What We Learned From 100 Million Open Source Downloads — Yuxuan Zhang, Z.ai
Overview
This talk introduces Z.ai's GLM 4.6 model series, highlighting its advancements in both language and multimodal understanding. The series has achieved over 100 million downloads across various platforms, indicating significant community adoption and contribution. GLM 4.6 demonstrates strong performance on public benchmarks, rivaling and sometimes surpassing commercial models, particularly in coding and reasoning tasks.
Who should watch
- AI engineers interested in the latest open-source model developments.
- Product Managers and builders evaluating models for integration.
- Developers seeking high-performance models for coding and reasoning tasks.
- Researchers exploring advancements in multimodal AI.
Key takeaways
- The GLM 4.6 series represents a significant leap in performance, outperforming its predecessor GLM 4.5 and competing strongly with commercial models like Claude 4 on benchmarks such as LMSYS Arena.
- Training for GLM 4.6 involved a multi-stage approach, starting with general pre-training on trillions of tokens, followed by reasoning, mid-training on code and project structures, synthetic reasoning data, and finally long context and agent data, pushing context windows up to 200,000 tokens.
- Z.ai has open-sourced its reinforcement learning framework, SLIDE, which supports both synchronous and decoupled training modes to optimize for different task types, from short reasoning to complex agent tasks.
- The GLM 4.5 V model offers enhanced multimodal capabilities, supporting both image and video understanding by preserving original input resolution and aspect ratios, and incorporating time index tokens for temporal understanding.
- The models are accessible through open-source weights via frameworks like Llama Factory and MS Swift, as well as through Z.ai's platform for direct use and integration.
- A new benchmark, CCBench, has been developed to test agent-style coding in real-world scenarios, evaluating models on tasks ranging from front-end development to data analysis and algorithm implementation.
- For scientific reasoning, data quality is emphasized over raw size, with a small set of expert-verified, high-quality multiple-choice questions yielding better performance than mixed-quality data.
Notable quotes
*GLM 4.6 shows a clear gain over GLM 4.5 and also outperforms open-source models released in the same period.*
*On LMSYS Arena, GLM 4.6 is tied for number one together with GPT-4 and Claude 4.5, and it's the only open-source model here.*
*For scientific reasoning data, quality really matters more than raw size.*
Unofficial community note. Prefer the recording for nuance.