LM Studio and Qwen3.8-27B: a serious AI that stays on your computer

My take: For private notes, an AI on my own machine beats any cloud. I'd still like LM Studio open source, like the engines under it.
What to know
- Alibaba's Qwen team released Qwen3.8-27B on Hugging Face on August 14, 2026, as an open-weight model: you download the model itself, under the Apache 2.0 license. It has 27 billion parameters (the values it learned, a rough measure of size) and reads text, images and video. On SWE-bench Pro, a coding test, it scores 61.7 against 53.4 for Claude Opus 4.6 Max: Qwen's own figures, not an independent test.
- LM Studio, the free app that runs it, asks for at least 16 GB of memory for the smallest version, a 16.1 GB file. That version is quantized: each parameter is stored on fewer bits, for a smaller file and a little less precision. Unsloth's guide, linked by Qwen, is less generous: it puts the 4-bit version on a Mac with 24 GB, and asks 31 GB at 8 bits.
- Once the model is on your disk, LM Studio runs without a connection, and its documentation says nothing you type in a chat leaves your device. The app is free but not open source; the engines under it, llama.cpp and MLX, are.
Spotted on @paul.automatise's reel, in French.
What changes
For what you would rather not hand to a cloud service, such as a contract or client notes, a model of this level fits on a well-equipped computer, with no subscription. The bill is the hardware: 16 GB is a floor that only takes the most compressed version, and on a Mac, LM Studio needs Apple Silicon. Watch the paid plans too: Bionic+, $20 a month, adds models hosted in the US, which run on LM Studio's servers, not yours. Check your memory, take the largest version that fits, and test it on a task you know well.
Source: Qwen Team, Alibaba, “Qwen3.8-27B”, August 14, 2026
Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.