Qwen’s coding update keeps a 27B footprint.

A downloadable dense model for coding and multimodal reasoning, with Apache 2.0 licensing.

Language & visionDownloadable

By Privaj Scout · Brief published Sep 10, 2026 · Sources checked Sep 10, 2026

SCOUT’S TAKE

Worth evaluating for self-hosted coding. The advertised improvement comes from Qwen’s own tests, not an independent Scout ranking.

See the comparison

Parameters: 27B
4-bit weights only: ≈ 13.5 GB
Native context: 262,144 tokens
License: Apache 2.0

What changed.

The April 21 launch post introduces the 27B dense model.
The model card lists 262,144 tokens of native context.
Qwen reports stronger coding results than Qwen3.5-27B.

Qwen launch post · April 21, 2026 ↗

SIDE BY SIDE

What Qwen’s tests report.

Maker-reported scores from one source. Higher is better on these tests; they do not establish an overall winner.

Qwen3.5-27B compared with Qwen3.6-27B
What matters	Qwen3.5-27B	Qwen3.6-27B
SWE-bench Verified	75.0	77.2
Terminal-Bench 2.0	41.6	59.3
Independent Scout test	Not performed	Not performed

Before you run it.

13.5 GB is theoretical weight storage: 27 billion × 4 bits ÷ 8. Runtime, image processing, quantization overhead and context cache require additional memory. This is not a measured VRAM requirement.

What the evidence shows

The figures below are copied as factual scores from Qwen’s model card and remain maker-reported. Match the evaluation setup before comparing them with other leaderboards.

Read another briefFLUX.2 [klein]