TAI #185: China's Open-Weight Holiday Blitz; GLM 4.7, Minimax M2.1 & MAI-UI

2024-09-10 · Source: Towards AI Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Intermediate, long

Summary

The AI landscape saw significant activity over the holiday period, particularly from Chinese and open-source developers, with major model releases and M&A deals. Z.ai launched GLM 4.7, a large-scale Mixture-of-Experts (MoE) model featuring "Interleaved Thinking" and "Preserved Thinking" for complex agentic workflows, achieving 85.7% on GPQA-Diamond and 42.8% on Humanity's Last Exam (HLE) with tools. Minimax introduced M2.1, a 229-billion-parameter MoE model optimized for software production and multi-language programming, scoring 88.6% on the VIBE benchmark. Alibaba's Tongyi group released MAI-UI, a family of GUI agents (2B to 235B parameters) designed for direct screen interaction using a device-cloud collaboration system, achieving a 76.7% success rate on AndroidWorld. Independent benchmarks from Artificial Analysis show GLM 4.7 leading open-weight models with an Intelligence Index score of 68, while Minimax M2.1 offers significantly lower inference costs at ~$128 compared to GLM 4.7's ~$334, though both exhibit a "grounding gap" with higher hallucination rates than proprietary models.

Key takeaway

For CTOs and VPs of Engineering evaluating AI model adoption, the emergence of high-performing open-weight models like GLM 4.7 and Minimax M2.1 presents a compelling cost-performance trade-off. While these models offer near-frontier reasoning capabilities at a fraction of the cost of proprietary APIs, you must implement robust verification systems to mitigate their higher hallucination rates. Consider a bifurcated strategy, leveraging "Thinker" models for complex reasoning and "Doer" models for high-volume, cost-sensitive execution to optimize your inference budget and application development.

Key insights

Open-weight models are closing the intelligence gap with proprietary models but show a widening safety gap in reliability.

Principles

• Agentic models benefit from persistent reasoning chains.
• Efficiency and cost drive application-layer model design.
• Direct GUI interaction simplifies agent deployment.

Method

GLM 4.7 uses "Interleaved Thinking" and "Preserved Thinking" to maintain reasoning chains. MAI-UI employs a device-cloud collaboration system to route tasks and reduce cloud calls for GUI agents.

In practice

• Route complex planning to "Thinker" models like GLM 4.7.
• Use "Doer" models like Minimax M2.1 for coding tasks.
• Explore GUI agents for automating legacy applications.

Topics

• Open-Weight Models
• AI Agents
• Model Benchmarking
• AI Hardware
• Mixture-of-Experts

Code references

• google/A2UI
• QwenLM/Qwen-Image-Layered
• facebookresearch/perception_models
• parcadei/Continuous-Claude-v2

Best for: Investor, CTO, VP of Engineering/Data, AI Engineer, AI Product Manager, AI Researcher

Related on AIssential

• The New Open-Weights Leader, Big AI’s Political Influence, Predicting Illness, Faster Reasoning — Open-weights AI models are rapidly closing the performance gap with proprietary frontier models, offering high-performance, customizable alternatives.
• AgentFloor: How Far Up the tool use Ladder Can Small Open-Weight Models Go? — Small open-weight models can handle most routine agentic tool use, reserving frontier models for complex long-horizon planning.
• How are Chinese models so strong with so little investment? — Chinese AI models achieve high performance through distillation from Western models and a robust national AI ecosystem.
• TAI #210: GLM-5.2 Closes Most of the Open-Weight Gap in Ten Weeks — GLM-5.2 demonstrates rapid open-weight model advancement through architectural, long-context, and reinforcement learning innovations, closing the gap with frontier models.
• not much happened today — The AI landscape is rapidly evolving with open-weight models nearing frontier capabilities, while agentic systems and specialized benchmarks drive practical application and efficiency.
• Open-Source AI Gains Ground as Rising Costs Push Shift to Smaller Models — Open-weight AI models offer substantial cost savings and near-parity performance compared to closed models, yet remain significantly underutilized.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI Newsletter.