Qwen 3.8 27B overthinks everything by default
Qwen 3.8 27B, a 27 billion parameter, Apache 2.0 licensed, vision-capable model from Alibaba's Qwen research lab, was released on a Friday, and the reviewer tested it two days later on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, running LM Studio's 17GB Q4_K_M quantized build and, separately, llama-server directly on the Spark. Qwen's own self-reported benchmarks show gains over the model's predecessor, Qwen 3.6 27B, and over the closed-weight Qwen 3.7-Plus, though the reviewer notes no independent benchmarks existed yet at the time of writing.
The headline problem is the model's default reasoning depth. Qwen's documentation sets the default reasoning_effort to xhigh, described as being for complex tasks demanding thorough analysis, versus medium for balancing accuracy and speed and low for optimizing speed and cost, and the LM Studio build preserves that default. The reviewer calls it a hilarious default and says it is absolutely not a good way to run the model on consumer hardware. A first attempt at generating an SVG of a pelican riding a bicycle, after raising LM Studio's context window from its default 8,192 tokens up to the model's full 262,144 token maximum (needed because the reasoning was eating the smaller window), took 21 minutes and burned 22,276 reasoning tokens to produce just 3,223 tokens of output. The same prompt with reasoning switched off took 137 seconds and produced 3,715 tokens. A far simpler prompt, asking only for an SVG of a circle, still ran at the default xhigh setting: the reasoning trace shows the model deciding to build a full 'geometric study' with concentric guide circles, tick marks, gradients and ambient animation rather than a plain circle, taking several minutes to deliver something the reviewer says was entirely not what had been asked for. The reviewer's recommendation is to ignore the xhigh default and run the model at low or no reasoning to start.
Outside the reasoning-effort problem, the reviewer found the model capable. Asked to return bounding boxes for pelicans in a photograph on a 0 to 1000 scale, it produced accurate coordinates. The reviewer then had the model build, from a single prompt and while running fully offline on the laptop, an HTML tool to visualize such bounding boxes over an uploaded image; without being asked, the model added a demo mode that draws its own placeholder pelican scene on canvas, reasoning in its trace that it needed a self-contained way to demo the tool without depending on an external image. Repeating that build prompt with reasoning off produced a version that almost worked but placed the boxes incorrectly, which the reviewer takes as evidence that reasoning did help here even though it was over-engineered. The reviewer also drove the coding agent Pi with the model, connecting it through LM Studio on the Spark shared over Tailscale, and got a solid answer to a question about how authentication works in a codebase, plus a working Python script, written by the model itself, to convert the resulting session transcript from JSONL into markdown.
On raw speed, the reviewer got roughly 15 to 30 tokens per second from LM Studio, which they call not terrible but slow enough to make it hard to prefer over hosted API models; for comparison, the reviewer cites Artificial Analysis figures of 74 tokens per second for OpenAI 5.6 Sol and 184 tokens per second for OpenAI 5.6 Luna. Qwen 3.8 27B supports Multi-Token Prediction, where a cheaper mechanism guesses several tokens ahead and the main model verifies the guesses, and following a tweet from llama.cpp creator Georgi Gerganov, the reviewer ran the model with MTP enabled via llama-server's draft-mtp spec-decoding mode on the Spark. A comparative benchmark, run with the help of GPT-5.6 in Codex, found that configuration outperformed the LM Studio default GGUF by about 72%. The reviewer expects further community speedups in the coming weeks. The closing take is that a 17GB file doing all of this on home hardware is remarkable progress for local models, but that performance, not capability, is what keeps it from being a daily driver: as a dense, non-Mixture-of-Experts model it needs heavy memory bandwidth, and neither machine the reviewer used is a top performer on that front.
Key facts
- Qwen 3.8 27B is a 27B parameter, Apache 2.0 licensed, vision-capable model from Alibaba's Qwen lab, distributed as a 17GB Q4_K_M quantized build.
- Its documented default reasoning_effort is xhigh; at that setting a pelican-on-a-bicycle SVG took 21 minutes and 22,276 reasoning tokens versus 137 seconds with reasoning off, and a request for a plain circle produced an unrequested animated 'geometric study' instead.
- The model handled 0-1000 scale bounding-box detection on pelican photos accurately and, in one prompt, built a working offline HTML tool to visualize those boxes, adding an unrequested placeholder demo scene.
- Driven through the coding agent Pi via LM Studio, it answered a codebase question about authentication and wrote a working Python script to convert a session transcript to markdown.
- Raw throughput was about 15-30 tokens/second versus 74-184 tokens/second cited for OpenAI 5.6 Sol/Luna; running llama.cpp's draft-mtp Multi-Token Prediction mode instead of the LM Studio default GGUF gave about a 72% speed boost in a comparative benchmark.
Why it matters
Qwen 3.8 27B is presented as evidence of how far local, open-weight models have come: a 17GB file running on a laptop or a DGX Spark that can handle vision tasks, tool calling and a coding agent loop, at a level the reviewer says would have been competitive with the best proprietary models a year ago. It also doubles as a cautionary example about defaults: a capable model shipped with a reasoning setting so aggressive that it makes trivial requests, like drawing a circle, take minutes and produce unrequested output, which can make the model look far worse than it is to anyone who does not know to change the setting.
Who it affects
Developers and hobbyists running LLMs locally on consumer hardware such as high-memory MacBooks or an NVIDIA DGX Spark, people evaluating open-weight Qwen models against closed-weight ones like Qwen 3.7-Plus or hosted APIs, and builders experimenting with local coding agents such as Pi or offline vision tooling.
How to use it
The model is available as a 17GB Q4_K_M quantized GGUF build usable through LM Studio or directly via llama-server, and supports a documented reasoning_effort parameter with three levels: xhigh (the default, for complex tasks), medium (balancing accuracy and speed) and low (optimized for speed and cost). The reviewer's strong recommendation is to ignore the xhigh default and start at low or no reasoning. LM Studio's default 8,192 token context limit is too small for the model's reasoning output and should be raised toward the full 262,144 token maximum. The model can also be pointed to from OpenAI-compatible tools, as shown by wiring it into the Pi coding agent over a Tailscale-shared LM Studio endpoint, and its speed can be substantially improved by serving it through llama-server with the --spec-type draft-mtp Multi-Token Prediction mode instead of the plain LM Studio default.
How solid is it
The account is a direct hands-on test across two machines with concrete, measured numbers: token counts and timings for the same prompts with reasoning on and off, working bounding-box output on a real photo, a working self-built visualization tool, a real Pi coding-agent session, and a benchmark comparing serving methods. The one caveat the reviewer flags is that Qwen's own headline benchmark comparisons against its predecessor and against Qwen 3.7-Plus are self-reported by Qwen, with no independent benchmark results available yet at the time of writing.
Risks and caveats
The default xhigh reasoning setting is described as a bad place to start: it can turn a request for a plain circle into an over-engineered animated illustration and turn a pelican SVG into a 21-minute, 22,000-plus-token job, which is likely to mislead anyone who runs the model out of the box. Even outside that failure mode, raw generation speed of 15-30 tokens per second is well below the 74-184 tokens per second cited for comparison hosted models, and as a dense, non-Mixture-of-Experts model it depends on memory bandwidth that neither of the reviewer's two machines has in abundance, so the faster serving setup involving Multi-Token Prediction requires extra configuration rather than working out of the box.
“My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.”
— the reviewer