How to Run Qwen 3.8 27B on Your Own Computer

Qwen 3.8 was the model everyone was waiting for and nobody could download. That changed in the middle of August 2026, when Alibaba put the weights on Hugging Face. Two of them matter to most people, and they did not arrive on the same terms. The 2.4 trillion parameter giant opened up on 12 August under Alibaba's own qwen3.8-max licence, and it needs a data center either way. Qwen 3.8 27B followed on 13 August under Apache 2.0, and it is a dense model built to fit on a single graphics card or a decent laptop.

This post is about the 27B. What it is, what it needs from your machine, and the one thing that trips up almost everyone on the first try.

What you are actually downloading

27 billion parameters, 64 layers, and a design that is unusual in a way that pays off later. Most models of this size run full attention on every single layer. Qwen 3.8 27B runs a cheaper linear attention on three layers out of four and saves the expensive kind for every fourth layer, so only 16 of its 64 layers do the heavy work. It reads text, and with a small extra file it reads images and video too.

The context window is 262,144 tokens out of the box, which is roughly a 600 page book in one sitting, and it stretches further with the usual scaling tricks.

Apache 2.0 is the licence, which is the permissive one. You can use it commercially, modify it, and ship it inside a product without asking anyone.

What it costs in disk and memory

Nobody runs the raw weights at home. You run a quantised version, which is the same model with the numbers stored more coarsely, and the community had these up within hours of the release. The sizes below are what Hugging Face reports for the unsloth/Qwen3.8-27B-GGUF pack, read on 14 August 2026. Different packs differ by a few hundred megabytes, so check the one you actually download.

Quant	File size	Roughly who it is for
UD-IQ2_XXS	9.0 GB	12 GB cards, quality takes a visible hit
UD-Q2_K_XL	10.7 GB	12 GB cards, almost no room for context
UD-Q3_K_XL	13.4 GB	16 GB cards with room for context
IQ4_XS	15.7 GB	the largest one that fits a 16 GB card whole
Q4_K_M	17.1 GB	the default everyone should start with
Q5_K_M	19.8 GB	24 GB cards, slightly sharper
Q6_K	22.9 GB	24 GB cards, close to lossless
Q8_0	29.0 GB	32 GB and up, effectively lossless
BF16	53.8 GB	the unquantised original, from the ggml-org pack
mmproj	0.9 GB	the extra file that gives it eyes

Q4_K_M at 17.1 GB is the sweet spot. On a 24 GB card it loads entirely into video memory with room left for context. A 16 GB card is the awkward size: Q4_K_M spills into system RAM, which costs speed but works, and IQ4_XS at 15.7 GB is the largest quant that stays on the card. On an Apple Silicon Mac with 32 GB of unified memory it is comfortable, and on 24 GB it is fine as long as you are not running much else. If you are unsure where your machine lands, our guide on how much RAM you need for local AI walks through it.

The context window is cheaper than you think

Here is where that unusual attention design shows up on your electricity bill. Every model keeps a running cache of what it has read, and that cache grows with every token. On a conventional 64 layer model of this shape it would cost about 256 KB per token, so a 100,000 token document would eat roughly 25 GB of memory on top of the model itself. That is why long context so often turns out to be theoretical.

Qwen 3.8 27B only pays for its 16 full attention layers, which works out to 64 KB per token. The same 100,000 token document costs about 6 GB. That is the difference between long context being a spec sheet number and long context being something you actually use on a Tuesday afternoon.

The one setting that decides everything

If you take one thing from this post, take this. Qwen 3.8 ships with its own chat template, and the template is not decoration. It is what tells the model where your message ends and its answer begins, and it is what switches its reasoning on and off.

Load the model without it and you get one of two failure modes, both of which look like a broken model rather than a broken setting. Either it rambles and never stops, or it answers in a strange clipped voice and forgets the conversation between turns. People report this as "the quant is bad" or "27B is overrated". It is neither.

In llama.cpp the fix is one flag, --jinja, which tells it to use the template shipped inside the GGUF file. In LM Studio and Ollama the template usually comes along automatically with the model, so this mostly bites people who downloaded a bare GGUF by hand. If your answers look wrong, check the template before you blame the weights. Our comparison of Ollama, LM Studio and LU Labs covers which tool handles this for you.

Giving it eyes

Vision lives in a separate file, the mmproj one in the table above, and it is under a gigabyte. Point your runtime at both files and the model can look at screenshots, photos, diagrams and video frames. Skip it and you have a very good text model that will politely tell you it cannot see the image you just pasted.

Thinking, and turning it off

Qwen 3.8 reasons before it answers. That is excellent for a hard question and wasteful for "what is the capital of Portugal", so the reasoning is a switch rather than a fixed trait. Turn it down for quick answers, turn it up when the question deserves it. On a local machine the difference is not money, it is your own time waiting for tokens, which for a lot of people is the more annoying currency.

If your machine says no

Not every laptop has 16 GB to spare, and the truly large Qwen 3.8 is out of reach for everyone outside a data center. That is the gap our cloud fills. Qwen 3.8 A95B, the 2.4 trillion parameter version with 95 billion parameters active per token, is in the LU Labs Cloud catalog on the Pro plan, alongside Kimi K3 and Ling 3.0 Flash. Same picker, same credits, no download.

The honest split is this. Run 27B at home when the material should not leave your machine, when you want it available offline, and when you would rather pay once for a graphics card than per token forever. Reach for A95B in the cloud when the problem is genuinely hard, or when the thing you are working on is bigger than your hardware. Most people who use both end up doing exactly that, and if you want the wider picture first, our roundup of the best open weight models you can run at home in 2026 puts this release in context.