Qwen 3.6 27B VRAM & GPU Requirements (2026)

Qwen 3.6 27B is about 16 GB of weights at Q4. Add room for context and you want roughly 20 GB to start, 24 GB to be comfortable. So a 24 GB card like a used RTX 3090 runs it at Q4 with headroom, a 16 GB card is limited to the smaller Q2 quant, and the table below shows exactly where your card lands. Once it fits, the real questions are how mature the software stack is, how fast it decodes, and what you pay to get there.

VRAM required to run Qwen 3.6 27B, by quantization

Weights are llama.cpp K-quant file sizes; min VRAM loads the weights plus a short context; comfortable adds headroom for a longer window. The KV cache grows with context. Size yours in the calculator below.

Can your GPU run Qwen 3.6 27B?

Fit means the weights plus a working context sit inside the card's usable VRAM at that quant. 16 GB cards run it at Q2 only; Q4 wants 24 GB or more. Apple and other unified-memory machines share VRAM with the OS, so keep a few GB in reserve.

Measured tokens per second for Qwen 3.6 27B

These are model-specific, single-stream decode numbers: first-party where we have it, community-reported otherwise. State the quant and the recipe, because a tuned run (speculative decoding, NVFP4, MTP) can nearly double a default one.

• RTX 5090: roughly 90 to 160 tokens per second, depending on the recipe.
• RTX 3090 (used): roughly 70 to 85 tok/s with the same tricks.
• R9700, B70 and MI50: slower at dense decode, in the order their bandwidth predicts.

Will it fit your card? VRAM calculator

Enter your VRAM, pick a context length and a quant, and we'll size Qwen 3.6 27B against it (weights + FP8 KV cache + runtime overhead).

Dense Qwen 3.6 27B vs the 35B-A3B MoE sibling

If decode speed on the cheaper cards matters more than peak quality, Qwen ships a sibling: Qwen 3.6 35B-A3B, a mixture-of-experts that activates only 3B parameters per token. It fits the same 32 GB at Q4 and feels snappier on bandwidth-limited cards. Want the 27B's coding quality, run the dense model; want speed on a budget card, the A3B MoE is the play. Our model picker shows both with per-card numbers.

Rated bandwidth vs what the software actually uses

Memory bandwidth is the ceiling, but you only reach it if the inference kernels keep the bus busy every cycle.

On the CUDA cards the kernels run the bus close to flat out, so decode lands roughly where the rated bandwidth predicts.

The younger stacks leave a big chunk on the table — the Arc Pro B70 saturates only about half its 608 GB/s on dense decode today.

These utilisation figures are rough and config-dependent — treat them as the shape of the gap, not a benchmark. The point holds: on immature stacks, raw bandwidth overstates real decode speed.

One more thing worth saying out loud, because it saves money: spending $2,700 to $4,700 on a big unified-memory box like a Strix Halo or DGX Spark does not unlock a meaningfully better coding model than this 27B on a 24 GB 3090. The dense 27B beats every open-weight MoE that fits in 128 GB or less. Capacity buys frontier MoE models that need 600 GB or more, not a better experience at the 27B tier.

The cheapest builds that run Qwen 3.6 27B

Popular single-card builds: Single RTX 5090 build · Single RTX 3090 (used) build · Single AMD Radeon AI Pro R9700 32 GB build · Single Intel Arc Pro B70 build · Single AMD Instinct MI50 32 GB (used) build