An open-source model family trained directly by Alibaba. It is one of the most widely used names in the open-model camp. Its flagship model is Qwen3.8, a 27B-parameter open-weight model released in August 2026 that supports multimodal input. That same month, it ranked among the top entries on Hugging Face's local inference model rankings. In August 2026 reporting on Anthropic's IPO, Qwen was also mentioned as investors pressed for details on the impact of Chinese AI models.
Current rank (1M)
19
Rank over the last 30 days
2026.08.20 – 2026.09.18 · High 10 · Low 19 · Now 19
Alibaba's official Wan account posted a clip from the eighth episode of its live sessions on September 17. Keith Zhang, the filmmaker behind Soulscape, said the 30-second single shot and an omni-reference that takes five videos surprised him most.
Google Research on September 15 published Retrieve-for-Train, a framework that returns search results as one coherent set, alongside its ICML 2026 paper. Running reinforcement learning once to teach a small diffusion model made query fan-out 12 to 20 times faster.
In a new test that hands AI a year of running an online store, GPT-5.6 Sol took first place while an older Alibaba model went bankrupt four times
When we connect two models, we've always had one write text for the other to read.
Two hours by train from Beijing, roughly 100 data centers have sprung up in Ulanqab, Inner Mongolia. Cheap electricity, cold winters, and surplus wind power have drawn China's AI companies to this plateau — but underneath sit two unresolved problems: water scarcity and coal. One city tells you how China's AI infrastructure math actually works.
A single round of fine-tuning was enough to strip the safety guardrails off 21 open-weight models.
Nick Tasker's 30-second ad "Let's Give Chicken a Break" went from a single base image straight through to final grading
Alibaba's specialized retraining for coding and collaboration reclaims the top spot on WorkArena, though terminal coding still lags
Z.ai's IPO gong, an order system rebuilt on Kimi K3, free DeepSeek access, and a 300-person WAIC party: founders, developers, and investors keep converging on this one spot, turning it into lived-in infrastructure for China's AI ecosystem.
The 125B MoE model runs on RAM alone with no GPU required, and it beat Claude Opus 4.6 Max across several benchmarks
Perplexity says its 27B on-device model paired with a custom harness outperformed open-source competitor harnesses on knowledge-work, web search, and document-parsing tests
Unsloth has lowered the barrier for fine-tuning a 27B model with free notebooks and optimized kernels.
