Qwen /

Qwen3.6-27B

A 27B open-weight model from Qwen that reads text and images, built mainly for coding work.

2.3k likes
2.4M downloads a month
28B parameters
Some experience
Vision & language

Open on Hugging Face (opens in a new tab)

Qwen3.6-27B is a language model with a built-in vision encoder, released as the first open-weight model in the Qwen3.6 series. It follows the Qwen3.5 series from February. The card says it was shaped by community feedback and aims at stability and practical use, especially for coding. The weights are in Hugging Face Transformers format and the card says they work with Transformers, vLLM, SGLang and KTransformers.

Two things the card highlights: better agentic coding, and a 'Thinking Preservation' option that keeps reasoning context from earlier messages so iterative work needs less repeating. It handles 262,144 tokens of context natively, and the card says that can be extended up to 1,010,000 tokens. On the card's own benchmark table it scores close to Claude 4.5 Opus on some tests and below it on others, for example 77.2 vs 80.9 on SWE-bench Verified. These are the Qwen team's numbers, with settings described in the card's footnotes, so treat them as a guide and not a promise for your own tasks. The excerpt I was given cuts off partway through the notes, so I can't say what the rest of the card covers, including any usage commands.

What you need to run it

28B parameters. The weights take about 56 GB at full 16-bit precision (fits a 128 GB Mac or a GPU with 128 GB), or about 17 GB as a 4-bit download, which runs on a 24 GB Mac or a GPU with 24 GB. Long inputs and images need more memory on top.

What it does well

Coding agent results
Reported scores include 77.2 on SWE-bench Verified, 53.5 on SWE-bench Pro and 59.3 on Terminal-Bench 2.0. The Terminal-Bench figure ties Claude 4.5 Opus (59.3) in the card's table.
Beats its predecessor and a larger sibling on some tasks
On SkillsBench it scores 48.2, against 27.2 for Qwen3.5-27B and 30.0 for the much larger Qwen3.5-397B-A17B. On SWE-bench Verified it also beats both of those (77.2 vs 75.0 and 76.2).
Strong on reasoning and maths
Reported scores include 87.8 on GPQA Diamond, 94.1 on AIME26 and 83.9 on LiveCodeBench v6, according to the card's table.
Very long context
262,144 tokens natively, extensible up to 1,010,000 tokens, which helps when you want to feed in a big codebase or long documents.
Thinking Preservation
An option to keep reasoning from earlier messages in the conversation, meant to make iterative development smoother and cut overhead.

What you could build

Ask it to read a whole repository and explain how the parts fit together.
Build a coding assistant or agent that edits files and runs terminal commands.
Generate frontend pages and web apps, including from a screenshot or design image, since it accepts image input.
Work through maths or science problems step by step and check its reasoning.
Run it behind your own server with vLLM or SGLang to experiment with an open model of your own.

Get started

The card says the weights work with Hugging Face Transformers, vLLM, SGLang and KTransformers, so pick one of those and load Qwen/Qwen3.6-27B. If you just want to chat with the Qwen models first, the card links to Qwen Chat. The part of the card I was given has no install or run commands, so check the full card and the Qwen blog post for those.

Commands are copied from the model card. Read it on Hugging Face for the full setup.