Qwen 3.8 35B-A3B leaks in GitHub commit, hinting at new MoE model
Alibaba's Qwen 3.8 35B-A3B spotted in a model training framework commit.
Deep Dive
A Reddit post titled “Just wait and see” points to a commit in the modelscope/ms-swift repository on GitHub, submitted by user BazzyIm. The article doesn’t provide any further details about what the commit contains.
Key Points
- New Qwen 3.8 35B-A3B model identified in a commit to Alibaba's ms-swift framework
- MoE architecture: 35B total parameters, 3B active per token for efficient inference
- Likely to run on 24GB VRAM GPUs, enabling local deployment and fine-tuning
Why It Matters
This leak signals a near-term release of a powerful, efficient open-weight model that lowers barriers for on-premise AI deployment.