Alibaba released Qwen 3.8, a new open-weight model family with two checkpoints: a 27.8B dense vision-language model that runs on a single 24GB consumer GPU, and a 2.4T-parameter (95B active) mixture-of-experts flagship, the first Qwen Max-class model with open weights. Benchmarks show the 27B model beating Claude Opus 4.6 Max on coding and computer-use tasks like SWE-bench Pro and OSWorld-Verified, though it lags on harder reasoning tests. The 2.4T flagship tops PaperBench versus GPT-5.6 Sol and Claude Opus 4.8, and QwenCloud pricing undercuts Anthropic and OpenAI significantly. Quantized GGUF builds let the 27B run on GPUs with as little as 12GB VRAM, and a 35B-A3B model is expected next though unconfirmed.
