On September 22, Liu Dayiheng, head of the Qwen LLM project at Token Foundry under Alibaba’s ATH Business Group, took the stage at the 2026 Apsara Conference in China with a talk titled “Qwen: Toward Real-World Agents.” But the big screen on stage displayed “Qwen4 Series Coming Soon,” with three names listed below: Qwen4-Max, Qwen4-Flash & Qwen4-Plus, and Qwen4-27B. This was the Qwen4 family’s first public appearance. No model launch was announced during the entire talk; Qwen4 is still in the training stage, and what Alibaba presented was a roadmap, not a downloadable model.
Max targets the capability ceiling, Flash and Plus focus on value for money, and 27B is reserved for local.
According to ZDNet’s on-site report, Liu Dayiheng’s Qwen4 plan is divided into three tracks: Max to push the capability ceiling, Flash and Plus to balance efficiency and price, and a 27B version for local deployment.
27B’s appearance on the Qwen4 list drew the most direct reaction from the community. On X, some commented, “As long as there’s a 27B, that’s enough,” while others asked whether the previous generation’s 35B was being discontinued. This size is anticipated because Qwen3.8-27B became the most-liked open-source large model in Hugging Face history on September 16, surpassing popular models such as FLUX.1, DeepSeek-R1, and Kimi-K3.
Qwen4 is still training; 5 trillion to 10 trillion is the goal for Qwen4.5 and Qwen5.
Liu Dayiheng called Scaling the path to ASI. He recalled that in the Qwen2.5 era, the flagship scale was 72B; today, Qwen3.8-Max has reached 2.4T total parameters, an increase of about 33 times in two years; Qwen4, based on a new architecture, has already entered training, and the subsequent Qwen4.5 and Qwen5 are planned to push the total parameter count to 5 trillion to 10 trillion.
Dayiheng Liu, born in 1993, was selected for Huawei’s “Genius Youth Program” in 2020 and moved to Alibaba DAMO Academy in 2021, where he led pretraining for the entire Qwen1 to Qwen3.5 series. In March this year, after former lead Junyang Lin left and Jingren Zhou became Chief Scientist, he took over coordination of Qwen. This was also his first time on the main forum of the Apsara Conference.
RSI’s external validation completed 33 rounds in one month.
At the conference, Alibaba laid out the results of Recursive Self-Improvement (RSI). Qwen3.8-Max can now autonomously set up training steps, construct training data, design experiments, and identify defects within the training process. It operated for more than a month with no human involvement, completed 33 effective iterations, and raised its Artificial Analysis index from 40 to 45, a 12.5% improvement.
The same mechanism has also been extended to inference and chip design. On T-Head’s new GPU, which it had never encountered before, Qwen3.8-Max independently tuned the inference framework for the next-generation Qwen3.8-Flash, increasing single-instance throughput by 96%. In an industrial-grade EDA environment, given only a real bus module specification, it ran autonomously for more than 60 hours and invoked tools over 10,000 times, ultimately reducing chip area by 42%, standard cell count by 29%, and power consumption by 59.5%.
New architecture open-sourced first, Qwen3.8-Flash training cost reduced to one-ninth
The official Qwen account open-sourced the Qwen3.8-Flash weights as early as August 26, and stated that it is a precursor version of the Qwen4 architecture: 125B parameters plus 51B N-gram embeddings, activating only about 6B per token, with training cost about one-ninth that of Qwen3.7-Plus, a native 262K context, extendable to 1M via YaRN.
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!
The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.
125B parameters + 51B N-gram… pic.twitter.com/SScnmzWS7O
— Qwen (@Alibaba_Qwen) August 26, 2026
The prices were also announced. Once the official version launches on the QwenCloud API, it will charge US$0.16 per million input tokens (about NT$5.2) and US$0.47 per million output tokens (about NT$15.3). Alibaba also presented its open-source results: it has open-sourced more than 460 Qwen models, with cumulative downloads exceeding 3 billion and more than 300,000 community-derived models. At the conference, it was also announced that in the past month alone, downloads of Qwen3.8-related models exceeded 56 million, with more than 1,900 derived models.
Omnimodal, speech, and video models updated in sync.
Beyond Qwen4, Alibaba also updated models for other modalities at the same conference. The omni-modal model Qwen3.8-Omni incorporates video, audio, images, and text into a single understanding framework. Alibaba says its overall cost is more than 90% lower than the previous generation. On-site demonstrations included automatically editing long-form footage into a finished video and generating an MV for music from a single sentence.
The Qwen-Audio-3.1 speech family is divided into three product lines: speech transcription, speech synthesis, and real-time interaction. There are also two new models based on a new architecture, Qwen-Audio-3.1-ASR-Next and TTS-Next; the simultaneous interpretation model Qwen3.8-LiveTranslate made its debut, reducing average per-character latency from 2.8 seconds to 2.3 seconds, while the average latency of human simultaneous interpretation is over 4 seconds. In video generation, Wan3.0 supports generating a 30-second video in a single pass and ranks first on both the Artificial Analysis text-to-video and video editing leaderboards. Alibaba ATH Technical Vice President Zheng Bo said on site that within three years there will be a native full-modal unified model, and the experience will no longer be limited by modal boundaries.
Conclusion
Putting 27B into the Qwen4 family shows Alibaba has not pulled back from its open-source path; writing 5 trillion to 10 trillion parameters into the timeline also brings the training-cost issue for next-generation models to the fore earlier. Next, watch what licensing terms Qwen4-27B offers and where the local hardware threshold falls; this determines whether it can replicate the local deployment momentum that Qwen3.8-27B sparked in the community.
Source: KOCPC Chinese