Qwen 3.8 27B Is Capable but Defaults to Extreme Overthinking

Alibaba's Qwen research lab released Qwen 3.8 27B on 16 August 2026 — an Apache 2.0-licensed, vision-capable model with 27 billion parameters that fits into a 17 GB Q4_K_M quantized file. Simon Willison tested it on a 128 GB M5 Max MacBook Pro and an NVIDIA DGX Spark using LM Studio and llama-server. The verdict: the model punches well above its weight class, but ships with a default reasoning configuration that will immediately frustrate anyone who doesn't read the docs.

The Overthinking Default

Qwen 3.8 27B exposes a reasoning_effort parameter with three tiers: xhigh (for complex tasks demanding thorough analysis), medium (balancing accuracy and speed), and low (optimising for speed and cost). The model defaults to xhigh. In practice, that setting consumed LM Studio's default 8,192-token context limit on trivial prompts. Expanding the context to the model's full 262,144-token maximum resolved truncation, but the cost became time: a single SVG-generation prompt consumed 22,276 reasoning tokens to produce 3,223 output tokens — a 21-minute wait. Asked to draw a circle, the model's reasoning trace spent several minutes deliberating palette options (deep teal ink, vermilion on off-white, a "Bauhaus/compass-drawing vibe") before producing a fully animated, multi-layered SVG bearing no resemblance to a simple circle. Reasoning off produced the same prompt's SVG in 137 seconds and 3,715 tokens. This directly affects prompt engineering decisions at the production level — the gap between xhigh and low is not a quality dial so much as a cost-explosion switch.

Vision and Bounding Box Accuracy

On vision tasks, the model delivered concrete results. Given a photograph and asked to return bounding boxes on a 0–1000 scale, it produced coordinates [195, 290, 370, 780] and [445, 320, 675, 850] for two pelicans — a close match when rendered over the original image. Willison then prompted the model to build an HTML bounding-box visualisation tool in a single turn. With reasoning enabled, the model built the requested interface and autonomously added a canvas-generated demo scene — procedurally drawing pelican silhouettes at the exact 0–1000 bbox positions — because the label string "pelicans" in the example JSON was apparently sufficient justification. With reasoning disabled, the tool ran but positioned boxes incorrectly, suggesting that for multi-step UI generation, some reasoning depth earns its keep even if xhigh is overkill.

Coding Agent Capability

The model was configured to drive the Pi coding agent via LM Studio on the DGX Spark, shared over Tailscale. Pointed at a Datasette codebase and asked "how does auth work?", it issued a sequence of tool calls across multiple files and returned a coherent summary. A follow-up prompt produced a working pi_jsonl_to_md.py script that converted the session's JSONL transcript to Markdown. Both tasks completed without human intervention mid-loop, validating the three prerequisites for agentic use — long context, code generation, and reliable tool-calling — at a weight size that remains accessible to capable small language models closing the gap on frontier agents.

Inference Speed and the MTP Optimisation

The most actionable speed gain comes from Multi-Token Prediction, an architecture feature baked into the model where a lighter draft mechanism guesses several tokens ahead and the main model verifies them. Invoking it via llama-server with --spec-type draft-mtp and --spec-default flags delivered approximately a 72% throughput improvement over the LM Studio baseline on the Spark. The local numbers still trail hosted APIs by a wide margin — 15–30 tokens per second versus 74 for OpenAI 5.6 Sol and 184 for 5.6 Luna. The gap is structural: dense (non-Mixture-of-Experts) models are memory-bandwidth-bound in a way that consumer and prosumer hardware cannot fully compensate for.

A 17 GB file delivering long-context reasoning, vision, tool-calling, and agentic code generation under an Apache 2.0 licence represents a capability tier that would have required proprietary frontier models a year ago. The xhigh default is a misconfiguration Qwen should fix, but with MTP optimisation already being explored two days post-release, inference performance is likely to improve further before most teams finish evaluating the model.