Qwen 3.6 Needs About 24GB to Run Fully on a GPU

To run Qwen 3.6 entirely in fast memory you need about 24GB for it: a 24GB graphics card, or a Mac with 32GB or more. The smaller model, Qwen3.6-27B, is a 17GB download at four bits in Ollama, and context comes on top of that; with less, the 35B-A3B can split between a smaller card and system RAM, more slowly.

The catch most people hit first is that ollama run qwen3.6 does not fetch that one. Ollama’s default tag shares its digest with qwen3.6:35b, a 23–24GB download.

Two models, and the bigger one is the default

Alibaba’s Qwen team released two open-weight Qwen 3.6 models in April 2026: the 35B-A3B on 14 April and the 27B a week later. Both read images as well as text, and both have a native context of 262,144 tokens.

	Qwen3.6-27B	Qwen3.6-35B-A3B
Design	Dense	Mixture of experts
Parameters, total / active per token	27B / 27B	35B / 3B
Native context	262,144 tokens	262,144 tokens
Ollama, 4-bit (q4_K_M)	17GB	24GB
Ollama, 8-bit (q8_0)	30GB	39GB
Ollama, full precision (bf16)	56GB	71GB
Hardware class at 4-bit, our reading	24GB GPU, or a Mac with 32GB+	32GB+ GPU, a Mac with 48GB+, or a smaller GPU with experts in system RAM

Sizes are Ollama’s, from its tags page on 5 October 2026. The last row is our arithmetic, not anyone’s specification: the file has to fit, then context and the operating system need room beyond it. Ollama lists no smaller Qwen 3.6 than these two, so it has no 8B-class Qwen 3.6 for a 12GB card.

“A3B” means less work, not less memory

The 35B-A3B is a mixture-of-experts model. Qwen’s card lists 256 experts, of which 8 routed and 1 shared are used for each token — about 3 billion parameters doing the work at any moment.

All 35 billion still have to be loaded. That is why its four-bit file is 24GB against the dense model’s 17GB, and why “only 3B active” does not make it the small one.

What it changes is where the weights can live. llama.cpp has a dedicated option, --cpu-moe, to “keep all Mixture of Experts (MoE) weights in the CPU”, and --n-cpu-moe to do it for some layers only. A dense model has no equivalent split, because every weight is used for every token. Our reading, unmeasured: a 12GB or 16GB card plus enough system RAM to hold the rest is a workable home for the 35B-A3B, and a slow one for the 27B. What happens when a model spills to system RAM covers how to see that split with ollama ps.

If the 27B fits, it is the stronger coding model by Qwen’s own numbers: 77.2 on SWE-bench Verified against 73.4 for the 35B-A3B, from Qwen’s model card and its own test setup.

Qwen asks for 128K of context. Ollama starts you far lower

Both model cards say the same thing: “we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.” Qwen’s own suggestion for an out-of-memory error is to reduce the window — and then it asks you not to go below 128K.

Ollama’s defaults, set by how much VRAM it finds:

• Under 24 GiB — 4k context
• 24–48 GiB — 32k
• 48 GiB or more — 256k

So on most home machines, Qwen 3.6 starts with a small fraction of the context its makers recommend. Raising it costs memory, and Ollama’s documentation says so directly. Our piece on why context, not the model, is the wall for local coding is the longer version of this argument, and it applies here with more force, because Qwen’s figure is roughly twice the 64,000 tokens Ollama recommends for coding tools.

One detail softens it. Qwen’s cards publish the layout: in the 27B, 16 of 64 layers use full attention and the rest use Gated DeltaNet, a linear-attention design; in the 35B-A3B it is 10 of 40. Our reading is that only those full-attention layers keep a cache that grows with every token, so long context should cost less memory here than in a model where every layer does. Neither Qwen nor Ollama publishes the figure, and we have not measured it.

On a Mac

Ollama publishes MLX builds for Apple silicon: 19GB for the 27B and 24GB for the 35B-A3B. A Mac’s GPU draws on unified memory, so the whole machine’s memory is the pool, as covered in what 512GB of unified memory changes. By our arithmetic, a 32GB Mac holds the 27B with about 13GB left for macOS, apps and context; the 35B-A3B on 32GB leaves 8GB, which is why we put it at 48GB.

What to buy, by what you have

• 16GB of VRAM or less: neither model fits on the card. Try the 35B-A3B with experts in system RAM, and have 32GB or more of system RAM so the 24GB file has somewhere to go.
• 24GB card: pull qwen3.6:27b, not the default, then raise the context deliberately and check ollama ps reads 100% GPU.
• 32GB+ card or a 48GB+ Mac: either model fits at four bits with room to raise context toward Qwen’s 128K.
• 64GB+: the 8-bit builds (30GB and 39GB) become practical. The rule of thumb for what fits on a card you already own works the same way for any size.

Both models carry the Apache 2.0 licence on their model cards — commercial use included, without the use restrictions in some licences that are open-weight but not open source.

Questions people ask

Can I run Qwen 3.6 on a 16GB GPU?

Not entirely on the card. The smallest four-bit build Ollama lists is the 27B at 17GB, more than 16GB before any context. It can still run with part of the model in system RAM, more slowly. Of the two, the 35B-A3B is the one built for that split.

How much VRAM does Qwen 3.6 27B need?

Ollama’s four-bit build is 17GB and the 8-bit build 30GB. Context comes on top, so in our reading a 24GB card is the smallest that holds the four-bit model with room to work.

Which Qwen 3.6 should I download, 27B or 35B-A3B?

The 27B if it fits your memory: it is 7GB smaller at four bits and scores higher on Qwen’s own coding-agent benchmarks. The 35B-A3B if it will not: only 3 billion of its parameters work per token, and llama.cpp can keep its experts in system RAM.

Can a Mac run Qwen 3.6?

Yes. Ollama publishes MLX builds for Apple silicon at 19GB (27B) and 24GB (35B-A3B). In our reading, 32GB of unified memory suits the 27B, and the 35B-A3B wants 48GB to leave room for macOS and context.

What context length should I use for Qwen 3.6?

Qwen advises at least 128K tokens to keep its thinking mode working well. Ollama gives far less by default below 48 GiB of VRAM — 4k under 24 GiB — so set it yourself, and expect memory use to rise when you do.

Is Qwen 3.6 free for commercial use?

Both open models carry the Apache 2.0 licence on their model cards, which allows commercial use, modification and redistribution, with no user-count ceiling of the kind some open-weight licences set.