Qwen 3.8 27B is excellent, but it defaults to overthinking things

TL;DR Highlight

Qwen 3.8 27B runs well locally, but its default xhigh reasoning is painfully slow

Who Should Read

Backend developers trying to run open-weight LLMs and coding agents on a laptop or local GPU

Core Mechanics

Default reasoning is xhigh, so it overthinks
- Qwen's docs set reasoning_effort to xhigh by default, and the LM Studio GGUF keeps it.
- The author says that is a poor default on consumer hardware.
- LM Studio's 8,192-token default context was used up by thinking, so he raised it to 262,144.
Pelican SVG took 21 min, 137s without reasoning
- At xhigh it used 22,276 reasoning tokens for 3,223 output tokens over 21 minutes.
- With reasoning off it produced 3,715 tokens in 137 seconds.
- The author says the wait was absolutely not worth it.
Asked for a circle, got an animated one
- For "draw an svg of a circle", xhigh thought for minutes and made an animated circle nobody asked for.
- He recommends starting at low or no reasoning.
Bounding boxes and coding agents work well
- Pelican boxes on a 0-1000 scale matched the photo well.
- Driving the Pi agent, it gave a solid answer about auth in the datasette codebase.

Community Reactions

Only three effort levels really work, per testing
- One commenter's comparison showed only none, low and xhigh behave differently, with low and medium nearly the same.
- Another said that counting 3090 electricity, Luna high via API is cheaper.
Thinking budget tames overthinking, per experience
- A commenter said llama.cpp's --thinking-budget and --thinking-message let you cut off and redirect the thinking.
- It can be controlled dynamically per message.
Plan first with none, then code, per experience
- A commenter set reasoning_effort to none, asked for a plan first, then asked for the code, and got reasonable code in reasonable time.
Too slow to use in practice, per experience
- One commenter's task took 11 hours on a dual GPU setup, though the result was the best of any self-hostable model.
- GPT 5.5 did a similar task in about 20 minutes.

How to Apply

Drop reasoning_effort to low and compare
- Raise the context length first, then start your local model at low or none.
- Run the same prompt at xhigh and low and record time and token counts.
Split reasoning level by task type
- Default to none or low for simple generation and Q&A.
- Raise it only for tool-building tasks that must be right in one shot, and check the difference.
Turn on MTP in llama.cpp and measure speed
- Start the server with a draft model and --spec-type draft-mtp.
- Compare tokens per second against your current server using the same prompt.

Try It

copy & run

# Paste into terminal: llama.cpp MTP launch from the article (llama serve command as written there)

llama serve \
  -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \
  -hfd ggml-org/Qwen3.8-27B-GGUF:Q4_0 \
  --spec-default \
  --spec-type draft-mtp \
  --reasoning-preserve

# Bounding box test (model loaded in LM Studio, llm CLI)
llm -a {{image URL}} \
  -m lmstudio/qwen/qwen3.8-27b \
  'Return JSON bounding boxes for the pelicans in this photo, 0-1000 scale for each dimension'

# Next: run the same prompt at reasoning low and xhigh and compare time and token counts

Terminology

reasoning_effortA dial for how long the model thinks before answering. Higher can be more accurate but is slower and uses more tokens.

Multi-Token Prediction (MTP)A cheap predictor guesses several tokens ahead and the main model verifies them in one pass. Like writing a draft first and only grading it quickly.

Q4_K_M quantizationShrinks model weights to about 4 bits to cut file size and memory. Like compressing a photo to save space.

Dense modelA model that uses all its parameters every step. Unlike Mixture-of-Experts, it needs lots of memory bandwidth.

Bounding boxRectangle coordinates around an object in an image. Used to draw boxes on photos.