Best GPU for Qwen 3.8 in 2026: Why 16GB Is Not Enough

Quick answer: Qwen 3.8 27B needs a 24GB card. The default Ollama build is an 18GB download at Q4_K_M, which does not fit in 16GB with room for context — the card that was comfortable for Qwen 3.6 is not comfortable here. A used RTX 3090 at roughly $820 is the cheapest card that holds it properly, and 32GB is what you want if you intend to use the 256K context or run it at Q8.

Qwen 3.8 at a glance

Alibaba released Qwen 3.8 in mid-August 2026. The headline change for local users is not the benchmark score — it is that the 27B is dense, not a mixture-of-experts, and it carries a vision encoder and a 256K context window.

Dense matters for your hardware budget. On an MoE model, total parameters set the VRAM floor while only the active ones cost you speed, which is why a large MoE can be tolerable on modest hardware. A dense 27B gives you no such discount: every parameter is read for every token.

VRAM requirements

These are the actual download sizes from Ollama’s model library, not estimates:

The 18GB figure is the whole story for most buyers. A 16GB card cannot hold an 18GB file, so the popular mid-range tier — RTX 4060 Ti 16GB, RTX 4070 Ti Super, RTX 5070 Ti — is out unless you accept a smaller quantisation or push layers to system RAM, which is slow enough to change how the model feels to use.

KV cache sits on top of every figure above, and it grows with context. Qwen 3.8 advertises 256K, but filling it is a memory decision rather than a free feature: a long-context session on a 24GB card will eat the headroom that the 18GB weights leave behind.

What the vision encoder costs

Qwen 3.8 takes image input as well as text, and the encoder is included in those download sizes. It is not free at runtime, though — images are turned into tokens, and a high-resolution image can consume a meaningful slice of your context window before the model has written anything.

If you only want text, the sizes above are what you plan around. If you intend to feed it screenshots or documents, treat the 24GB tier as the floor rather than the target.

How fast will it be?

Token generation on a local GPU is bound by memory bandwidth far more than by compute: the card reads the whole weight set once per token, so an 18GB model on a 936 GB/s RTX 3090 is doing substantially more memory work per token than a 7B model on the same card.

We publish modelled figures rather than measurements, and the tokens per second estimator explains the basis. What is worth knowing before you buy is the shape: moving from a 4GB 7B model to an 18GB 27B model multiplies the per-token memory traffic by roughly four and a half, and no amount of compute makes up for that.

Which GPU should you buy?

• Cheapest card that actually fits? A used RTX 3090 at around $820. 24GB, 936 GB/s, and the same capacity as a 4090 for roughly a third of the price. See our used RTX 3090 buying guide before you buy one.
• Buying new and want it to last? The RTX 4090 at ~$2,200. Same 24GB, considerably faster, and a warranty.
• Want Q8 or the long context? The RTX 5090 at ~$4,900 is the only consumer card with 32GB, which is what the 30GB Q8 build and heavy context use both want.
• Already own a 16GB card? Do not buy a second GPU for this. Run Qwen 3.6 or a 14B instead — both fit properly in 16GB, and a model that fits beats a larger one that spills.

Who this is for

This guide is for people choosing hardware around Qwen 3.8 specifically. If you are picking a GPU for local LLM work in general, the how much VRAM for local LLM guide is the better starting point, because the right answer there depends on the largest model you intend to run rather than on this one.

Common mistakes

• Assuming 16GB still covers the flagship 27B. It did for Qwen 3.6. An 18GB file in a 16GB card does not fit, and no quantisation trick changes that arithmetic — you either drop to a lower quant or accept offloading.
• Budgeting for the weights and forgetting the context. 18GB of weights in a 24GB card leaves about 6GB, and a long session will use it. The 256K number in the spec sheet is a ceiling, not an allowance.
• Treating the vision encoder as free. Images become tokens. A few high-resolution inputs can consume more context than a long conversation.
• Buying a 32GB card for a 7B workload. If you mostly run smaller models, the 27B is not a reason to spend $4,900 — the extra capacity sits idle.

Final verdict

Qwen 3.8 27B is the first Qwen release in a while that changes the hardware answer rather than just the benchmark table. The 18GB Q4_K_M download moves the entry point from 16GB to 24GB, and that single number decides most purchases: a used RTX 3090 at roughly $820 is the cheapest honest way in, an RTX 4090 is the same capacity done faster, and the 32GB RTX 5090 is the only card that makes Q8 or serious long-context use comfortable.

If your budget stops at 16GB, the right move is a smaller model rather than a bigger card — running Qwen 3.6 well beats running Qwen 3.8 badly.

Frequently asked questions

Can I run Qwen 3.8 on a 16GB GPU?

Not the default build. The Q4_K_M download is 18GB and a 16GB card cannot hold it, so you would need a smaller quantisation or offloading to system RAM, which slows generation considerably. A 24GB card is the realistic entry point.

How much VRAM does Qwen 3.8 27B need?

Roughly 18GB for the default Q4_K_M build, about 30GB at Q8, and around 56GB at BF16, with KV cache on top of each. In practice that means a 24GB card for everyday use and 32GB if you want Q8 or long context.

Is Qwen 3.8 worth upgrading your GPU for?

Only if you were already near the 24GB tier. If you own a 16GB card that runs Qwen 3.6 comfortably, the sensible path is to stay on a model that fits rather than spend four figures to hold a slightly better one.