Qwen Developers Reveal Upcoming 27B Model and Video Understanding Architecture in AMA
Qwen developers revealed during a Twitter/X AMA that they will soon release a highly capable 27B model (Qwen 3.8 27B) in response to community feedback. They also shared details on their 100-hour video understanding capability, which uses a hierarchical video memory system to encode segments into structured textual graphs. The release of a new 27B model fills a highly anticipated gap in the open-source AI ecosystem, offering a balance between size and performance. Additionally, their hierarchical memory approach to long-context video understanding could set a new standard for processing massive multimodal datasets without relying on complex agent swarms. The upcoming Qwen 3.8 model features 2.4T total parameters with 95B active parameters, and its architecture remains similar to Qwen 3.5 but scaled up. The developers also noted that they used an "unreasonable amount of compute" for reinforcement learning (RL) in post-training, though they will not release a technical report immediately to maintain their monthly release cadence.
## BACKGROUND
Qwen is a series of open-source large language models developed by Alibaba's team, known for competitive performance across various benchmarks. In AI, video understanding typically requires processing sequential frames, which researchers often tackle using either agent swarms (multiple AI agents collaborating) or semantic graphs that map out relationships between entities and events over time.