Qwen Devs AMA reveals 27B model, 100-hour video memory

During a Twitter/X AMA, the Qwen team behind Alibaba's open-weight models answered community questions with big reveals. Most notably, they confirmed a 27B model is releasing 'very soon,' directly addressing a skipped version from the previous cycle. They clarified it's not a retrain of the 3.6 27B but a new capability level. They also gave a peek at the upcoming Qwen3.8 architecture, which is similar to 3.5 but at a much larger scale: 2.4T total parameters with 95B active, using a hybrid attention approach. Training involved 'a truly unreasonable amount of compute' for reinforcement learning, and the team emphasized a solid data foundation. No technical report is planned for this release, as they maintain a near-monthly cadence, and more powerful models are already in the works.

Another standout was a response on 100-hour video understanding, describing a 'hierarchical video memory system' rather than an agent swarm. Video segments are encoded into a structured textual graph containing scenes, entities, events, and temporal relationships, enabling retrieval and reasoning across over 100 hours. The team also hinted at SAE-guided fine-tuning as a research technique that may inform future training. Community-driven release decisions were highlighted: the 27B revival reflects user feedback. For developers, updates to Qoder and QwenWork coding tools are coming soon, but no CLI interface was announced. The AMA suggests Qwen is aggressively pushing scale and reasoning, with practical implications for anyone building open-source AI applications.