Open Models Recap: Kimi K3, Qwen, and the Accelerating Open-Closed Gap

A quarterly roundup of open-weight AI models reveals accelerating progress from Chinese labs, shifting distillation dynamics, and new US ecosystem competitors.

MiHiR SEN
MiHiR SEN
·1 min read
This quarterly recap examines the accelerating progress of open-weight AI models, particularly from Chinese labs like Kimi and Zhipu. It challenges common assumptions about model distillation, noting that reinforcement learning remains difficult to distill effectively. The analysis also highlights emerging US competitors focusing on fine-tunability and domain-specific utility rather than raw benchmark chasing.

Acceleration in Chinese Open Models

The latest quarterly recap of open-weight AI models shows rapid progress from Chinese labs. Models like Kimi K3 and GLM 5.2 are demonstrating strong performance in agentic coding and computer use tasks. These labs appear to benefit from high capital efficiency, focused research teams, and increasing access to domestic compute resources, allowing them to close the gap with closed-source frontier models.

The Distillation Debate

A recurring topic is the impact of model distillation. While some argue that extracting reasoning traces from closed models gives open labs an unfair advantage, the reality is more nuanced. Distillation provides a boost during the supervised fine-tuning stage, but the core capabilities are forged during large-scale reinforcement learning. Generating high-quality RL data remains a difficult, unsolved research challenge that cannot be easily bypassed by simply copying API outputs.

US Ecosystem Adaptation

In response, US open-model players are adapting. Companies like Thinking Machines and Nvidia are focusing on releasing models that serve as excellent bases for domain-specific fine-tuning. Rather than chasing raw benchmark supremacy, these labs are building utility and developer mindshare by providing highly optimizable, transparent architectures.