Qwen3.8 27B (low)

Qwen 3.8 27B is an open-weights, multimodal large language model developed by Alibaba. Released as a dense model under the Apache 2.0 license, it is designed to run on consumer-grade hardware while delivering performance comparable to much larger proprietary models in coding and reasoning tasks. It is the open-weight counterpart to the much larger API-only Qwen 3.8-Max and is a direct successor to the Qwen 3.6 27B model.

Architecture and Capabilities

The model features a native multimodal architecture, allowing it to process text, image, and video inputs without relying on external vision adapters. Its underlying architecture utilizes a hybrid approach with 64 layers, combining standard gated attention with Gated DeltaNet, a linear-attention design that significantly reduces the memory footprint of the Key-Value (KV) cache. This efficiency allows the model to maintain a native context window of 262,144 tokens, which can be extended up to 1,000,000 tokens using YaRN scaling.

Reasoning and Performance

A notable feature of Qwen 3.8 27B is its adjustable reasoning_effort setting (low, medium, and xhigh). This allows users to control the depth of the model's internal "thinking" process. While higher settings improve performance on complex engineering and software development tasks—scoring 61.7% on the SWE-bench Pro—they can lead to significant generation latency and "overthinking" on simpler queries. Practitioners typically recommend adjusting the reasoning level to "low" for routine tasks to optimize throughput on local workstations.