Should I wait for Qwen 3.8 27B or run Qwen 3.6 now?

Run Qwen 3.6 now. The 3.8 27B weights are promised for the coming week, but GGUFs, tooling support, and abliterated builds all take additional time. Qwen 3.6 27B is proven, runs from 8 GB VRAM up, and switching models later is one click, not a migration.

Does Qwen 3.8 have a bigger context window than Qwen 3.6?

Yes: 1 million tokens for Qwen 3.8 Max versus 256K for Qwen 3.6. Whether the open weight 3.8 27B keeps the full 1M window is not yet confirmed.

Qwen 3.8 vs Qwen 3.6: What Actually Changed, and Which to Use

Q: Is Qwen 3.8 better than Qwen 3.6?

The Max is, by a wide margin on the launch benchmarks: it is a 2.4 trillion parameter frontier model, while Qwen 3.6 tops out at 35B. But they are not really competitors. Qwen 3.8 Max is a hosted API model today; Qwen 3.6 is the generation you can download and run on your own GPU right now.

Q: Is there an uncensored Qwen 3.8?

Not yet. Uncensored community builds need public weights, which are promised for the coming week. Qwen 3.6 abliterated builds exist today and are among the strongest uncensored local models of 2026.

With Qwen 3.8 Max generally available since August 3 and its open weights promised for the coming week, the practical question is not "which is better" (the 2.4T flagship, obviously) but "which Qwen should I actually be using, today and next month". Here is the honest lineup.

The Two Generations at a Glance

Feature	Qwen 3.6	Qwen 3.8
Released	April 21, 2026	August 3, 2026 (GA)
Sizes	27B dense, 35B MoE (3B active)	Max: 2.4T MoE (~95B active); 27B announced
Context window	256K	1M (Max; 27B unconfirmed)
Modalities	Text + image input	Text + image + video input
Open weights	Yes, since April	Promised for the coming week (Max + 27B)
Runs locally	Yes, from 8 GB VRAM	27B: soon. Max: never on consumer hardware
Uncensored builds	Yes, abliterated and heretic variants	Not until the weights are public
API price	Cheap everywhere (open weights, many hosts)	$2 / M input, $6 / M output, single upstream

Where 3.8 Is Simply a Different Class

Qwen 3.8 Max is a frontier model. Its launch scores (92.6 GPQA Diamond, 86.6 Terminal-Bench 2.1, 73.5 FrontierSWE) are not numbers a 27B or 35B produces; the agentic coding jumps in particular roughly double its predecessor generation. It also sees video, holds a million tokens, and ships built in tools. If your job is huge context analysis or hard agent runs and hosted is acceptable, the Max is the strongest Qwen ever offered. The access guide covers every route.

Where 3.6 Still Wins

It exists on your GPU. Qwen 3.6 27B runs today from 8 GB VRAM up, fully offline, no logging, no rate limits, free. The Max never will, and the 3.8 27B is still a promise plus a GGUF pipeline away. The setup guide takes about 15 minutes.
Uncensored builds exist. The abliterated and heretic Qwen 3.6 variants are proven, downloadable, and among the strongest entries on our 2026 uncensored models list. Qwen 3.8 has none until the weights drop, and then the community needs weeks.
Proven, not promised. 3.6 has three months of independent testing behind it. 3.8 has launch day screenshots.
Total privacy. A local 3.6 never phones home. The hosted Max is an API with logs, whatever the quality gap.

The Interesting Case: Qwen 3.8 27B

The announced open weight 27B is where the generations actually meet. If it inherits even part of the Max's agentic gains at 3.6 27B hardware cost (~16 GB at Q4), it becomes the default local Qwen the day usable GGUFs appear, and the abliterated builds that follow become the new uncensored baseline. The realistic timeline: weights within the week, GGUFs days later, tooling and abliterations over the following weeks. Our hardware breakdown tracks what is confirmed versus projected.

So Which Should You Use?

Your situation	Use
Private chat, writing, coding on your own GPU	Qwen 3.6 27B, today
Uncensored local AI	Qwen 3.6 abliterated, today
Million token context or hardest agent tasks, hosted OK	Qwen 3.8 Max via API
Want the newest local Qwen	Run 3.6 now, switch to 3.8 27B when GGUFs land (one click, not a migration)

FAQ

Is Qwen 3.8 better than Qwen 3.6?

The Max is, by a class margin, on launch benchmarks. But it is hosted only today, while 3.6 runs on your own hardware. Different tools for different jobs.

Should I wait for Qwen 3.8 27B?

No. Run 3.6 now; switching later is one click in the Model Manager. Waiting gains you nothing.

Does Qwen 3.8 have a bigger context window?

Yes, 1M versus 256K, for the Max. Whether the 27B keeps the full window is unconfirmed.

Is there an uncensored Qwen 3.8?

Not yet; that needs public weights. Qwen 3.6 abliterated is today's answer.

Getting Started

Whichever generation wins your workload, the local side takes five minutes to set up:

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
# Windows: setup.bat | Linux: ./setup.sh

Or grab the installer from the releases page, open the Model Manager, and search Qwen.

Locally Uncensored is AGPL-3.0 licensed and free to use. Built by PurpleDoubleD.