Latest open artifacts (#19): Qwen 3.5, GLM 5, MiniMax 2.5 — Chinese labs' latest push of the frontier

2026-03-03 · Source: Interconnects AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Data Science & Analytics · Depth: Intermediate, short

Summary

The latest intelligence brief highlights a busy month for open-weights AI, with new flagship model releases from Qwen, MiniMax, Z.ai, Ant Ling, and StepFun, alongside anticipation for DeepSeek V4. The report introduces "Relative Adoption Metrics (RAM)" to normalize model downloads, identifying GPT-OSS as a top performer and noting DeepSeek V3.2's underperformance compared to earlier 2025 releases. Featured models include Qwen 3.5 (0.8B to 397B-A17B, multi-modal, Qwen-Next architecture with GDN layers), Step-3.5-Flash (196B-A11B MoE, strong in math), GLM-5 (744B-A40B, leading to price hikes), and MiniMax-M2.5 (rivaling larger models). OpenThinker-Agent-v1, Tri-21B-Think, and MiniCPM-SALA are also noted for their specific capabilities.

Key takeaway

For AI Scientists evaluating new open-weight models, utilize the Relative Adoption Metrics (RAM) framework to assess true adoption and identify high-potential models like GPT-OSS. Pay close attention to the early adoption of models with hybrid architectures, such as Qwen 3.5 dense models, as their performance relative to established brands can indicate future trends. Consider the specific strengths of models like Step-3.5-Flash for specialized tasks.

Key insights

New open-weights AI models are rapidly emerging, with adoption tracked by Relative Adoption Metrics (RAM).

Principles

• RAM scores >1 indicate a model is on track to be a top 10 all-time downloaded model in its size class.
• Hybrid model architectures can push the limits of open-source tools.

Method

Relative Adoption Metrics (RAM) normalizes model downloads against peer models in their size class to gauge adoption and identify high-performing or underrated models.

In practice

• Consider Qwen 3.5 models for multilingual and general tasks, disabling reasoning for smaller models if overthinking occurs.
• Evaluate Step-3.5-Flash for math-intensive benchmarks.
• Explore OpenThinker-Agent-v1 for agentic reasoning tasks and evaluation.

Topics

• Open-weights AI Models
• Relative Adoption Metrics
• Qwen 3.5
• Mixture-of-Experts
• Agentic Reasoning

Best for: NLP Engineer, AI Scientist, Research Scientist, Machine Learning Engineer, AI Engineer, AI Product Manager

Related on AIssential

• Alibaba just open-sourced a model that rivals GPT-5.2 — Alibaba's Qwen 3.5, with 17 billion active parameters, rivals top closed-source models while being open-weight.
• Z.ai GLM-5: New SOTA Open Weights LLM — Chinese AI labs are rapidly advancing open-weight models, emphasizing cost-efficiency, long-context, and agentic capabilities.
• How Open-Weight Models Changed the AI Landscape — Open-weight models drive rapid AI innovation through shared designs and iterative improvements.
• Qwen3.5-397B-A17B: the smallest Open-Opus class, very efficient model — Qwen3.5 offers multimodal, long-context capabilities with efficient sparse MoE architecture, enabling local deployment despite its large scale.
• 2025 Open Models Year in Review — Open models significantly advanced in 2025, rivaling closed-source performance and fostering a more open ecosystem.
• MiniMax M2.7 matches GPT-5.3-Codex in software engineering tasks — Chinese AI labs are increasingly releasing large language model weights, though licensing terms vary significantly.

Open in AIssential →

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