Qwen 3.8 27B: Powerful Open-Weight LLM Overthinks, Runs Locally

Simon Willison's review of Qwen 3.8 27B, released on 16 August 2026 by Alibaba's Qwen research lab, highlights its impressive capabilities but notes a default 'extra high' reasoning effort that leads to excessive token consumption. The model, a 27B parameter vision-capable LLM under Apache 2 license, shows benchmark gains over its predecessor Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus. Willison tested it on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark using LM Studio's 17GB Q4_K_M quantized build. He found that the default reasoning effort caused the model to use up the 8,192 token context limit on simple tasks, but increasing to the full 262,144 context resolved this. A prompt for a pelican SVG took 21 minutes and 22,276 reasoning tokens, while turning reasoning off produced a result in 137 seconds. The model also excelled at bounding box detection and built a custom HTML tool from a single prompt. Willison tested it as a coding agent with Pi, successfully converting a JSONL transcript to markdown. Performance is a catch: 15-30 tokens per second from LM Studio, slower than hosted APIs. However, Multi-Token Prediction (MTP) support, as suggested by llama.cpp creator Georgi Gerganov, boosted performance by ~72% on the Spark. The model's ability to run on consumer hardware in a 17GB file demonstrates significant progress in local LLMs, making powerful AI accessible without expensive datacenter hardware.