What should I run locally until the Qwen 3.8 weights drop?

Qwen 3.6 is the best Qwen you can run today: the 27B dense runs on 8 to 16 GB of VRAM depending on quant, and its abliterated builds are among the strongest uncensored local models of 2026.

Can You Run Qwen 3.8 Locally? The Honest Hardware Math

Q: Can you run Qwen 3.8 Max on a normal PC?

No. At roughly 2.4 trillion parameters, a 4 bit quantization of Qwen 3.8 Max is around 1.2 terabytes of weights before any KV cache. The largest consumer machines top out around 512 GB of unified memory, which is less than half of what the weights alone need.

Q: Will there be a Qwen 3.8 GGUF?

Yes, and the one that matters is the 27B. The weights came out on August 13, 2026 under Apache 2.0 and community GGUFs appeared the same day. Q4_K_M measures 17.1 GB, which fits a 26 GB GPU.

Q: What hardware does Qwen 3.8 27B need?

Measured from the released GGUFs: Q4_K_M is 17.1 GB and wants a 24 GB card, IQ4_XS at 15.7 GB is the largest quant that stays whole on 16 GB, UD-Q3_K_XL is 13.4 GB, and the 2 bit quants sit at 9.0 to 10.7 GB for 12 GB cards at a visible quality cost.

Q: Will there be an uncensored Qwen 3.8?

Almost certainly, once the weights are public. Every open Qwen generation has received abliterated and heretic community builds within weeks. Hosted only models can never be freed; open weights are what makes uncensored variants possible.

Update, August 14, 2026: the wait is over. The 2.4T Max checkpoint went public on Hugging Face on August 12 as Qwen/Qwen3.8-2.4T-A95B under Alibaba's own qwen3.8-max terms, which is not Apache 2.0, and the consumer sized Qwen 3.8 27B followed on August 13 under Apache 2.0. Real GGUF sizes, the VRAM math and the chat template trap are in How to Run Qwen 3.8 27B Locally. Everything below is the pre drop picture, kept for context.

Alibaba's Qwen 3.8 Max went generally available on August 3 with the promise local AI people care about: open weights within the week. The question everyone is typing into a search box right now is whether a 2.4 trillion parameter model can run on hardware a human being owns.

(If you just want to use it today rather than host it, see How to Run Qwen 3.8 Max, with API options and working code.)

We love running models locally; it is the whole point of this site. So here is the answer with actual numbers instead of hype, and, unusually for a frontier launch, this one comes with genuinely good local news attached.

The Short Answer

The Max: no. Not on a gaming PC, not on a 4090, not on a maxed out Mac Studio. It is roughly 35 times larger than the biggest models people comfortably run at home.

The 27B: yes, very soon. The same open weight drop includes a Qwen 3.8 27B, and a 27B is exactly the size class that runs on a 16 GB GPU at 4 bit today. That is the model this page is really about.

The Math on the Max, Step by Step

Model memory needs come down to parameters times bytes per parameter. Qwen 3.8 Max has about 2.4 trillion parameters. At the standard quantization levels, the weights alone come to:

Precision	Bytes / param	Weights size	Runs on
FP8 (native serving)	1	~2.4 TB	Multi node GPU cluster
4 bit (Q4, the local standard)	~0.5	~1.2 TB	Still a cluster
2 bit (extreme, quality suffers)	~0.25	~600 GB	Two or more linked 512 GB machines

For scale: the largest single box a consumer can buy in 2026, a 512 GB unified memory Mac Studio, holds less than half of the Q4 weights. And that is before the KV cache. The headline 1 million token context costs additional memory that grows with every token in the window; fill even a fraction of it and you add hundreds of gigabytes on top.

The MoE architecture does not rescue you. Only ~95B of the 2.4T parameters compute per token, which saves processing power, but all of them have to sit in memory, because you never know which experts the next token will activate.

Someone will chain enough Mac Studios together to produce single digit tokens per second within weeks of the weights dropping. It will make a fantastic video, and it will change nothing about what you should run at home.

Qwen 3.8 27B: The Drop That Actually Lands on Your Desk

This is the difference between this launch and the Kimi K3 or DeepSeek V4 Pro launches: Alibaba is shipping a consumer sized model in the same open weight drop. What we can responsibly project for the 27B, based on the Qwen 3.6 27B precedent:

Your VRAM	Expected quant	Precedent (Qwen 3.6 27B)
8 GB	2 bit class (UD-IQ2_XXS)	8.7 GB on disk, reduced quality
12 GB	Q3_K_M class	13 GB, the RTX 3060 sweet spot
16 GB	Q4_K_M, the recommended default	~16 GB
24 GB	Q6_K, near lossless	~21 GB

The exact sizes are in now, measured rather than projected. From unsloth/Qwen3.8-27B-GGUF on August 14, 2026: UD-Q2_K_XL 10.7 GB, UD-Q3_K_XL 13.4 GB, IQ4_XS 15.7 GB, Q4_K_M 17.1 GB, Q5_K_M 19.8 GB, Q6_K 22.9 GB, Q8_0 29.0 GB, plus a 0.9 GB mmproj file if you want it to read images. The table above guessed 16 GB for Q4, which turned out a little optimistic: on a 16 GB card, IQ4_XS is the one that stays whole. Full breakdown in How to Run Qwen 3.8 27B Locally.

And weights mean derivatives. Every open Qwen generation has received abliterated and heretic builds within weeks, and those builds are what most of the best uncensored local models of 2026 are made of. A hosted only model can never be freed; an open 27B from the current frontier generation is about the best thing that can happen to this ecosystem.

What to Run Today

The 3.8 weights have landed, so on a 16 GB card or better the answer is now Qwen 3.8 27B. Below that, Qwen 3.6 is still the right call and is genuinely strong.

If you are not sure which tier your machine sits in, LU Labs breaks down how much RAM local AI needs at 8, 16, 32 and 64 GB.

Your VRAM	Run this	What you get
6 GB	Llama 3.1 8B Abliterated	Fast all round chat and coding, zero refusals
8 to 16 GB	Qwen 3.6 27B (abliterated variants exist)	Sharp reasoning; see the Qwen 3.6 local guide
16 to 24 GB	Gemma 4 27B class, 30B MoEs	Genuinely strong writing and analysis
48 GB+	70B class	The best private AI that currently fits on a desk

Locally Uncensored makes the whole thing one click: it auto detects 12 local backends, and its Model Manager downloads any of these without a terminal. New to this? Start with the 5 minute beginner guide.

And When You Really Need Max Class Quality

Be honest about the split: most everyday AI work runs beautifully on a local model with total privacy. For the rest (huge context jobs, the hardest agent tasks), use hosted access with open eyes about where your data goes. The API guide covers every route that works today, and once the Max weights are public and independently hostable, it becomes a candidate for neutral providers too, including LU Labs Cloud, the hosted studio from the team behind this app.

FAQ

Can you run Qwen 3.8 Max on a normal PC?

No. The weights alone are ~1.2 TB at 4 bit quantization; the biggest consumer machines hold 512 GB. It is a datacenter model.

Will there be a Qwen 3.8 GGUF?

Yes, and the one that matters is the 27B from the same drop: ~16 GB at Q4_K_M, which fits a 16 GB GPU. Max GGUFs will exist too, at ~1.2 TB, as homelab exotica.

What hardware does Qwen 3.8 27B need?

Projected from the 3.6 precedent: 16 GB VRAM for Q4_K_M, 12 GB for Q3 class, 8 GB at 2 bit with reduced quality. Confirmed numbers once GGUFs exist.

What should I run until then?

Qwen 3.6, which runs today from 8 GB up, including abliterated builds.

Will there be an uncensored Qwen 3.8?

Almost certainly, within weeks of the weights: every open Qwen generation has been abliterated. We will cover the builds when they exist.

Getting Started

Skip the 1.2 terabyte daydream and have a private, uncensored model running in the next five minutes:

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

Or download the installer from the releases page, open the Model Manager, and pick a model that fits your VRAM from the table above.

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