At the Apsara Conference on September 22, Alibaba unveiled its model family progress: Qwen3.8-Max launched to take No.1 on Artificial Analysis Agentic and CodeArena, lifting the AA global score from 40 to 45; Qwen4 with an all-new architecture is now in training, with Qwen4.5 and Qwen5 in the pipeline and total parameters heading toward 50-100 trillion.
RSI: the model iterating itself
The headline revelation is Recursive Self-Improvement entering training, inference and chip-model co-design: Qwen3.8-Max autonomously built training pipelines, constructed data, designed experiments and located defects for over a month — 33 effective iterations with zero human involvement. On inference, Qwen3.8 self-adapted the SGlang framework to an unseen new Pekochip GPU for +96% single-instance throughput; on chip co-design, it ran 60+ hours and 10,000+ EDA tool calls to achieve -42% area and -59.5% power.
Ecosystem: the world's most popular open family
Qwen3.8 models topped 56M downloads and 1,900 derivatives within a month; Qwen3.8-27B passed DeepSeek-R1 and Llama 3.1 to become the most popular model in Hugging Face history. Cumulatively: 460+ open models, 3B+ downloads. Airbnb's CEO says the company leans heavily on Qwen — "better and cheaper than OpenAI"; Pinterest built its AI shopping assistant on Qwen at under 8% the cost of closed peers; Reuters develops its own models on Qwen.
The multimodal matrix
- Video: Wan 3.0 — 30-second generation, structured input, native audio-visual sync — tops both AA video leaderboards; next-gen in training;
- Audio: Qwen-Audio-3.1 leads domestically across ASR/TTS/Realtime; LiveTranslate cuts average latency to 2.3s vs ~4s for human interpreters;
- Image: Qwen-Image-3.1 for second-level e-commerce design; the 7B Qwen-Image-2.1 fully open-sourced with native transparency;
- World model: HappyOyster 2.0 Preview debuts.
Qwen3.8-Flash open-sources the next-gen architecture early: training cost down ~90%, cached input at RMB 0.1 per million tokens.
(Facts aggregated from public reporting)