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

Qwen 3.8 27B, a 17GB local LLM, impresses with its vision, coding, and long-context abilities, rivaling models a year its senior. However, its default "xhigh" reasoning leads to delightfully over-engineered, albeit painstakingly slow, outputs. This model showcases incredible progress in local AI, albeit with the caveat that patience (or prompt engineering) is a virtue.

Score

Comments

Highest Rank

11h

on Front Page

First Seen

Aug 17, 1:00 AM

Last Seen

Aug 17, 11:00 AM

Rank Over Time

The Lowdown

Qwen 3.8 27B, a new Apache 2 licensed, vision-capable 27-billion parameter LLM from Alibaba, is making waves for its impressive performance on consumer hardware. Despite its relatively small 17GB footprint, the model demonstrates advanced capabilities that were once exclusive to much larger, often proprietary, models.

The model's default xhigh reasoning effort setting, while enabling thorough analysis, leads to significant "overthinking." This results in extremely slow response times (e.g., 21 minutes for an SVG generation) and highly verbose internal monologues, often producing hilariously over-engineered outputs for simple prompts like "draw an SVG of a circle."
Despite the speed penalty, the xhigh reasoning can yield superior results in complex tasks, such as generating an HTML tool for bounding box visualization, which was accurate even if unnecessarily elaborate, whereas a non-reasoning attempt failed.
Qwen 3.8 27B proves capable of driving coding agents like Pi, successfully analyzing codebases and generating accurate Python scripts for data transformation, further demonstrating its practical utility.
Performance remains the primary bottleneck; the model feels slow even without xhigh reasoning, typically achieving 15-30 tokens/second on consumer hardware. However, optimizations like Multi-Token Prediction (MTP) using llama.cpp can significantly boost inference speed.
The article concludes by highlighting the "miracle" of having such a capable, open-weight model runnable locally, emphasizing the rapid progress in the field and the potential for these models to perform real work without requiring massive datacenter resources.

The Gossip

Overthinking Odes and Operational Fixes

The default "xhigh" reasoning in Qwen 3.8 27B is a central topic, with users sharing experiences of its extreme verbosity and slowness. Many discuss strategies to mitigate this, including setting explicit reasoning levels (low, medium) and using prompt engineering techniques or `llama.cpp`'s `--thinking-budget` to curtail its elaborate internal processes. Some even propose creating proxies to interrupt over-analysis, turning the model into a more efficient agent.

Performance Ponderings and Pelican Prowess

While impressed by Qwen's advanced capabilities—such as generating high-quality pelican SVGs, accurate bounding boxes, and driving coding agents—commenters frequently highlight the significant performance trade-offs. The model's slowness is a major point of contention, with some tasks taking hours on dual-GPU setups, far exceeding the time taken by larger, hosted models. This leads to a discussion on the balance between deep reasoning and the practical need for speed in real-world applications, especially for agentic workflows where token efficiency directly impacts operational costs.

Local LLM's Leaps and Limits

The Hacker News community expresses considerable enthusiasm and "amazement" at the rapid progress in local LLMs. Qwen 3.8 27B is lauded for bringing high-end capabilities—like vision, coding agent support, and long context—to consumer hardware, effectively rivaling proprietary models from just a year prior. This is seen as a significant achievement, proving that powerful AI doesn't always require datacenter-grade hardware. However, the discussions also acknowledge the practical limits, particularly memory bandwidth constraints on current hardware (e.g., Apple's M-series chips), which remain a bottleneck for achieving optimal performance.