On August 16, 2026, Alibaba’s Qwen research lab released Qwen 3.8 27B, an Apache 2‑licensed 27‑billion‑parameter language model with vision capabilities.

The model was evaluated on a 128 GB M5 Max MacBook Pro and an NVIDIA DGX Spark using the 17 GB Q4_K_M quantized build in LM Studio and llama‑server. Its default reasoning setting (xhigh) caused the model to consume the full 262,144‑token context for simple prompts, resulting in extended generation times. A test generating an SVG of a pelican riding a bicycle with the full context took 21 minutes and used 22,276 reasoning tokens.

Running the same prompt with reasoning disabled reduced the time to just over two minutes and produced a shorter output. Independent benchmark comparisons indicated OpenAI’s 5.6 Sol at 74 tokens per second and 5.6 Luna at 184 tokens per second, showing that Qwen 3.8 27B operates at lower speeds on consumer hardware. The developer community has begun exploring speed improvements, including Multi‑Token Prediction support, which can increase inference throughput by roughly 72 % compared to the default GGUF deployment. The 17 GB file demonstrates that a 27‑billion‑parameter open‑source model can run on high‑end consumer machines, though performance remains a limitation for real‑time applications.