Which Qwen3 Size Fits Your GPU? 8GB to 24GB Cards

Qwen3 comes in eight open-weight sizes, and the one you can run comes down to a single number: how much memory your graphics card has. Alibaba's Qwen team released six dense models (0.6B, 1.7B, 4B, 8B, 14B and 32B) and two mixture-of-experts models (30B-A3B and 235B-A22B), all under the Apache 2.0 license. Picking the right one takes about a minute of arithmetic.

The formula

The rough fit check I use is VRAM (GB) = parameters in billions × bytes per parameter × 1.2 + 2. At Q4_K_M, the most common 4-bit GGUF quant, a parameter takes about half a byte. At 8-bit it takes one byte. The 1.2 covers the KV cache and runtime overhead at a modest 4K to 8K context, and the extra 2 GB is held back for the operating system and the inference engine itself.

Qwen3 sizes by card

Here is what that formula gives at Q4_K_M:

• Qwen3-4B: about 4.4 GB. Fits any 8 GB card, and the 8-bit version fits too.
• Qwen3-8B: about 6.8 GB. Fits an 8 GB card at a short context.
• Qwen3-14B: about 10.4 GB. Wants a 12 GB or 16 GB card.
• Qwen3-32B: about 21.2 GB. Needs a 24 GB card such as an RTX 3090 or RTX 4090.
• Qwen3-30B-A3B: about 20.0 GB. Also a 24 GB card.
• Qwen3-235B-A22B: about 143 GB. Not a single-GPU model at home.

The two smallest models, 0.6B and 1.7B, run on almost anything, including a laptop with no dedicated GPU. They suit quick jobs like sorting text or short autocomplete, and they run out of depth on long answers.

The MoE trap

Qwen3-30B-A3B has 30 billion parameters in total, but only about 3 billion are active for each token. That makes it quick for its size. It does not make it small. Every expert still has to sit in memory so the router can pick from them, so you size it on the 30B total, not the 3B active. A 16 GB card that handles Qwen3-14B comfortably will not hold 30B-A3B at Q4. On a 24 GB card the two big options need about the same memory, and the MoE version will usually generate faster because each token touches fewer weights.

Context is the hidden cost

The Qwen team lists a 32K context length for the 0.6B, 1.7B and 4B models and 128K for the 8B, 14B and 32B models. That long window costs memory, and the formula's 1.2 factor only assumes a few thousand tokens. Qwen3-32B has 64 layers, 8 key-value heads and a head size of 128, so a full 32K-token window adds roughly 8.6 GB of KV cache at 16-bit precision. On a 24 GB card already holding 21.2 GB of weights, that does not fit. Keep the window modest, or shrink the cache: Ollama's FAQ says an 8-bit cache (q8_0) uses about half the memory of the default f16, and cache quantization needs flash attention turned on.

Quick picks

On an 8 GB card, start with Qwen3-8B at Q4. On 12 GB, Qwen3-14B at Q4 fits with a little room for context. On 16 GB, run Qwen3-14B and spend the spare memory on a longer window, or run the 8B at 8-bit, which comes to about 11.6 GB. On 24 GB, Qwen3-32B at Q4 is the largest dense option, and 30B-A3B is the faster alternative at nearly the same memory.

In Ollama the tags are qwen3:8b, qwen3:14b, qwen3:30b and qwen3:32b. After the model loads, run ollama ps and read the PROCESSOR column. Anything under 100% GPU means part of the model spilled into system RAM, and generation slows down hard.

The full Qwen3 table, including 8-bit numbers for every size, is in the Qwen3 hardware requirements guide. To try a different context length or quant, the VRAM calculator runs the same math.

Sources: Qwen3 announcement (qwenlm.github.io/blog/qwen3), Qwen3-32B config on Hugging Face, Ollama FAQ (docs.ollama.com/faq). Memory figures are formula estimates, not benchmarks.