Qwen3.6's dense 27B is the no-fuss 24GB-card option — every parameter fires, nothing to offload, a real 256K context. The VRAM math, the GPU, and how it differs from the 35B-A3B MoE.
This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
Qwen3.6 comes in two shapes, and it's worth knowing which one you're getting. The default (ollama run qwen3.6) is the 35B-A3B mixture-of-experts model. This guide is about the other one: the dense 27B, the model to reach for when you'd rather have every parameter working on every token, with nothing to offload and no surprises.
What Qwen3.6 27B actually is
From the Qwen3.6 27B model card (read 26 Sep 2026): 27B parameters, a 262,144-token (256K) context natively (extensible further), Apache-2.0. It's dense — no experts. Its config.json shows Qwen's efficiency trick intact: most layers (48 of 64) use linear attention, with 16 full-attention layers, so the 256K context stays cheap on VRAM. (Our VRAM engine sizes the full model at about 27.8B, the figure the pages below use.)
Dense 27B vs the 35B-A3B MoE
This is the choice most people are actually making, so here it is plainly. The dense 27B activates all 27B parameters per token: predictable, no offload, and it fits a 24GB card whole at Q4. The 35B-A3B MoE stores 35B but activates only ~3B, so it's faster to generate — but it needs ~21 GB at Q4 (still 24GB territory) and its weights are a 35B's weights. If you want simple and predictable on 24GB, take the dense 27B; if you want more speed and don't mind the larger download, take the MoE.
At ~17.5 GB, Q4_K_M sits comfortably on a 24GB card with plenty of room for context — and thanks to the linear-attention layers, pushing toward that 256K window barely moves the number. Q6 (~23 GB) still fits a 24GB card; Q8 needs more. A 16GB card is too tight for Q4 here (17.5 GB), so this is genuinely a 24GB model.
On the quant labels: Q4_K_M compresses each weight to about four bits — near-full quality at a quarter of the size, and the build Ollama pulls by default — while Q5 and Q6 spend more VRAM for a little more fidelity, and both still fit a 24GB card here. For a dense 27B on 24GB the practical choice is Q4 for maximum context room or Q6 for the sharpest output, and the linear-attention design means you can lean on the long context either way without your VRAM climbing.
What Qwen3.6 27B is genuinely good for is the everyday heavy lifting where you'd rather have a predictable dense model than juggle a mixture of experts: long-document analysis, coding across several files, and general assistant work, all on one 24GB card under a licence (Apache-2.0) that lets you build on it commercially. It isn't the flashiest model in the Qwen line — the MoE variants and the giant Qwen3-Next get more attention — but for a lot of people it's the one that simply works without fuss.
A 24GB RTX 3090 was listing from about $0.139/hr on Vast.ai's spot market (26 Sep 2026) — enough for Qwen3.6 27B at Q4 or Q5. Prices move; check the live figure before you rent.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
How to run it
ollama run qwen3.6:27b # dense; the bare `qwen3.6` pulls the 35B-A3B MoE
Name the size: qwen3.6:27b gives the dense model, while the bare qwen3.6 tag resolves to the 35B-A3B mixture-of-experts. Both are worth having, but they're different models — this one is the simple, all-dense 24GB choice.
Who should skip it
Skip the dense 27B if you want the fastest generation on your 24GB card — the 35B-A3B MoE activates only 3B parameters and feels quicker. Skip it if you only have a 16GB card; it's too tight at Q4, so drop to Qwen3.5 9B or a smaller Qwen. And if you need image input, this is a text model — look at Gemma 4 31B instead.
Qwen3.6 27B: common questions
What's the difference between Qwen3.6 27B and 35B-A3B?
27B is dense — all parameters active, nothing to offload, fits a 24GB card whole at Q4. 35B-A3B is a mixture-of-experts: it stores 35B but activates ~3B, so it's faster but larger to store.
How much VRAM does Qwen3.6 27B need?
About 17.5 GB at Q4_K_M (computed by our VRAM engine, 16K context), which fits a 24GB card comfortably. A 16GB card is too tight for Q4.
What does 'ollama run qwen3.6' download?
The 35B-A3B MoE, not this dense 27B. For the dense model, run ollama run qwen3.6:27b.
Does the long context cost a lot of VRAM?
No. Most of Qwen3.6's layers use linear attention, so the 256K context barely adds to the figures above.
For the MoE alternative see Qwen3.6 35B-A3B; to compare 24GB-class models head to head, our which 24GB model to run guide puts it against Mistral Small 24B and Gemma 4 31B.
