Skip to main content

Qwen4 Is Now Training: Alibaba Maps Out Its Trillion-Parameter Ambitions

Alibaba Pushes the Boundaries of AI Scale

At the 2026 CloudCon in Hangzhou on September 22, Alibaba didn't just talk about the future—it laid out a concrete roadmap. The headline? Qwen4, built on a next-generation architecture, has officially entered training. And the company isn't stopping there. Plans are already in motion for Qwen4.5 and Qwen5, which will balloon to between 500 billion and 1 trillion parameters.

But the real story might be what Alibaba calls Recursive Self-Improvement (RSI). This isn't just about bigger models; it's about models that can improve themselves. Liu Dayiheng, who leads the Qwen LLM project, framed scaling as a critical path toward artificial superintelligence (ASI). And Alibaba is putting its money where its mouth is: after Qwen3.8-Max launched, it didn't just score well on benchmarks like Artificial Analysis Agentic agents and CodeArena front-end programming—it actually built its own training processes, constructed data, designed experiments, and identified defects entirely on its own for over a month. No humans involved. That autonomous loop completed 33 effective iterations and boosted its Artificial Analysis score by 12.5%.

More Than Just a Bigger Brain

Architecture matters, too. Qwen3.8-Flash slashes training costs by nearly 90% thanks to innovations in attention mechanisms, and it boosts inference efficiency by activating fewer parameters. Meanwhile, the multimodal lineup is getting a serious upgrade. Qwen3.8-Omni-Flash now handles video, audio, images, and text in one unified system. Wan3.0 can generate a 30-second video in one go and supports structured input—and a next-gen video model is slated for November.

On the audio front, the Qwen-Audio-3.1 series covers ASR, TTS, and Realtime, while Qwen3.8-LiveTranslate cuts average latency to just 2.3 seconds per character. And that's not all: Qwen-Image-3.1, Qwen-Image-2.1, HappyShrimp 1.1, and HappyOyster 2.0-Preview are all moving forward simultaneously.

Open Source by the Numbers

Alibaba's open-source ecosystem is thriving. In just one month, Qwen3.8-related models racked up over 56 million downloads and spawned more than 1,900 derivative models. To date, the company has open-sourced over 460 Qwen models, with total downloads exceeding 3 billion and over 300,000 derivatives. Big names like Perplexity, Airbnb, Pinterest, and Reuters are already using Qwen for agents, AI assistants, and custom models.

So what does this mean for the rest of us? If Alibaba's roadmap holds, the next few years could see AI systems that not only learn but improve themselves—and they'll be available to anyone with an internet connection.

Key Points:

  • Qwen4 is in training; Qwen4.5 and Qwen5 will scale to 500B–1T parameters.
  • Recursive Self-Improvement (RSI) is now active in training, inference, and chip collaboration.
  • Qwen3.8-Flash cuts training costs by ~90% and improves inference efficiency.
  • Multimodal upgrades include unified video/audio/image/text processing and a 30-second video generator.
  • Open-source Qwen models have surpassed 3 billion downloads and 300,000 derivatives.