The New Open-Weights Leader, Big AI’s Political Influence, Predicting Illness, Faster Reasoning

2026-02-20 · Source: The Batch | DeepLearning.AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, long

Summary

Z.ai has released GLM-5, a 744 billion-parameter Mixture-of-Experts transformer model with 40 billion active parameters per token, designed for long-running agentic tasks. It achieved top performance among open-weights models on Artificial Analysis' Intelligence Index, 𝜏²-Bench Telecom, Vending Bench 2, and Chatbot Arena Code, nearly matching proprietary leaders like Claude Opus 4.6 and GPT-5.2. GLM-5 supports up to 200,000 input tokens and 128,000 output tokens, features function calling, reasoning, and context caching, and is available via a free web interface, Hugging Face (MIT license), and API. The model was pretrained on 28.5 trillion tokens and utilized Z.ai's open-source reinforcement learning software, "slime," for post-training, along with DeepSeek sparse attention for efficient long-context processing. Additionally, Liquid AI introduced LFM2.5-1.2B-Thinking, a 1.17 billion-parameter hybrid transformer-convolutional neural network model optimized for on-device reasoning, operating under 900MB RAM and twice as fast as Qwen3-1.7B on CPUs. This model, pretrained on 28 trillion tokens, excels in agentic tasks and data extraction, supporting eight languages, but shows higher hallucination rates in knowledge-intensive tasks. Separately, researchers developed SleepFM, an AI system using CNN, transformer, and LSTM architectures to classify over 130 diseases, including heart disease and psychiatric disorders, up to six years before symptom onset, based on one night of sleep study recordings. SleepFM achieved higher AUC scores across various disease categories compared to systems without pretraining. Finally, top tech and AI companies, including Meta, Amazon, Alphabet, and Microsoft, spent over $100 million on lobbying in 2025, influencing policies like the reversal of the advanced AI chip ban to China and limiting state-level AI regulation, while also supporting large data center projects.

Key takeaway

For CTOs and VPs of Engineering evaluating AI model adoption, the emergence of high-performing open-weights models like GLM-5 and efficient edge models like LFM2.5-1.2B-Thinking presents compelling alternatives to proprietary solutions. You should explore these options for cost-effectiveness, customization, and deployment flexibility, especially for agentic workflows or on-device applications. Be mindful of the evolving regulatory landscape influenced by tech lobbying, which may streamline compute infrastructure but could also favor larger players.

Key insights

Open-weights AI models are rapidly closing the performance gap with proprietary frontier models, offering high-performance, customizable alternatives.

Principles

• AI tools can unleash human creativity and generate new job categories.
• Sparse attention and hybrid architectures improve on-device model efficiency.
• Sleep signal analysis can predict disease onset years in advance.

Method

GLM-5 uses a Mixture-of-Experts transformer architecture, pretrained on 28.5 trillion tokens, with post-training via "slime" reinforcement learning and DeepSeek sparse attention. LFM2.5-1.2B-Thinking employs a phased training approach for its hybrid transformer-convolutional architecture, including step-by-step reasoning data and merging specialized RL models. SleepFM combines CNN, transformer, and LSTM in a two-stage training process for sleep data encoding and disease classification.

In practice

• Utilize GLM-5 for complex, long-running agentic tasks requiring high token limits.
• Deploy LFM2.5-1.2B-Thinking for on-device agents, data extraction, or RAG applications.
• Consider sleep study data for early disease prediction using AI models.

Topics

• AI Job Creation
• Large Language Models
• AI Policy and Lobbying
• On-Device AI
• Medical AI

Code references

• zou-group/sleepfm-clinical

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

Related on AIssential

• TAI #185: China's Open-Weight Holiday Blitz; GLM 4.7, Minimax M2.1 & MAI-UI — Open-weight models are closing the intelligence gap with proprietary models but show a widening safety gap in reliability.
• Leadership Challenges at OpenAI Revealed — AI is achieving human-level performance in complex tasks while economic and regulatory challenges emerge.
• AI News: Everyone's Mad At Anthropic Now — Open-source AI models are rapidly closing the performance gap with proprietary models while offering significant cost and privacy advantages.
• The Sequence Radar #837: Last Week in AI: From Model Releases to Market Structure — AI is evolving into an infrastructure play, prioritizing scale, integration, and open models over raw product demos.
• What’s next for Chinese open-source AI — Chinese open-source AI models offer near-frontier capabilities at a fraction of the cost, rapidly reshaping global AI innovation.
• Why AI Users Are Raving About GLM 5.2 — GLM 5.2's strong performance in specific domains like web design challenges frontier model dominance and diversifies AI strategy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Batch | DeepLearning.AI.