Qwen VRAM Requirements: Sizes from 8B to 72B Compared
Qwen VRAM requirements depend on which model you pick, but the pattern is steady: at 4-bit, plan on roughly 0.6 GB per billion parameters for weights, then add the KV cache and 10% for buffers. Qwen3 8B needs about 7 GB for one user at an 8K context, Qwen3 32B about 24 GB, and the older Qwen2.5 72B about 55 GB.
The table below uses Qwen3 for the sizes up to 32B and Qwen2.5 for 72B, since Qwen3 has no dense model at that size. Newer Qwen releases may have changed since this was written, so the last section shows how to check any model yourself in two minutes.
Memory by model
Weights are the approximate Q4_K_M and Q8_0 GGUF sizes. The totals add the KV cache for one user at an 8K-token context and 10% overhead:
| Model | Q4_K_M weights | Q8_0 weights | KV cache per token | Total, Q4 | Total, Q8 |
|---|---|---|---|---|---|
| Qwen3 8B | ~5 GB | ~8.7 GB | 144 KB | ~7 GB | ~11 GB |
| Qwen3 14B | ~9 GB | ~15.7 GB | 160 KB | ~11 GB | ~19 GB |
| Qwen3 32B | ~20 GB | ~35 GB | 256 KB | ~24 GB | ~41 GB |
| Qwen3 30B-A3B | ~18.6 GB | ~32 GB | 96 KB | ~21 GB | ~36 GB |
| Qwen2.5 72B | ~47 GB | ~77 GB | 320 KB | ~55 GB | ~88 GB |
Reading across: a 24 GB card runs Qwen3 32B at Q4_K_M for one user with little margin. A 48 GB card runs it at Q8_0, or runs Qwen2.5 72B only at Q3 or below. An 80 GB card fits Qwen2.5 72B at Q4_K_M with room for around eight users at 8K.
Why 30B-A3B needs 30B of memory
Qwen3 30B-A3B is a mixture-of-experts model: 30.5 billion parameters in total, about 3 billion active for each token. The active count sets the speed. The total count sets the memory, because the server cannot know in advance which experts a request will use, so all of them stay loaded.
That makes it an attractive trade when you have the memory and want speed. It generates tokens at roughly the pace of a small model while answering closer to a mid-sized one. It is a poor choice when memory is the constraint: Qwen3 14B uses half the memory and, for many tasks, performs comparably. The same logic applies to the very large Qwen3 235B-A22B, whose weights are about 140 GB even at 4-bit, however few parameters are active.
Context length changes the answer
Qwen3 models handle 32K tokens natively and can be stretched to 128K with YaRN scaling, which Qwen documents on each model card. The KV cache scales linearly with that window:
| Model | 8K context | 32K context |
|---|---|---|
| Qwen3 8B | ~1.1 GB | ~4.5 GB |
| Qwen3 14B | ~1.25 GB | ~5 GB |
| Qwen3 32B | ~2 GB | ~8 GB |
| Qwen2.5 72B | ~2.5 GB | ~10 GB |
Qwen3 also has a thinking mode that writes out its reasoning before the answer. Those reasoning tokens occupy the cache like any others, so a thinking-mode request can use several thousand extra tokens of context. If you enable it, size for 16K or 32K even when your prompts are short.
Multiply by concurrent users: ten people on Qwen3 32B at 32K need about 80 GB of cache, four times the weights at 4-bit. Our GPU memory for LLM guide explains why concurrency, not the model file, is usually what runs a server out of memory.
Check any Qwen model yourself
The KV cache needs three numbers from the model's config.json: layers, key-value heads and head size. Fetch them and apply the formula:
curl -s https://huggingface.co/Qwen/Qwen3-8B/raw/main/config.json \
| python3 -c "import json,sys; c=json.load(sys.stdin); print(c['num_hidden_layers'], c['num_key_value_heads'], c.get('head_dim'))"
KV bytes per token = 2 x layers x KV heads x head size x 2 (bytes at 16-bit)
For Qwen3 8B that gives 2 x 36 x 8 x 128 x 2 = 147,456 bytes, or 144 KB. If head_dim is missing from a config, divide hidden_size by num_attention_heads. Replace the model name in the URL to check any other release.
Licenses
Qwen3 models are released under Apache 2.0, which allows commercial use without extra terms. Qwen2.5 is mixed: most sizes are Apache 2.0, but the 3B and 72B models use different Qwen licenses with their own conditions. Read the license on the model card of the exact size you plan to deploy. The open-source LLM for enterprise post has a short checklist for this.
Which plan size a Qwen model falls into
AI Refinery, a self-hosted LLM platform, sizes its plans by model parameters: Basic covers up to 14B, Professional up to 70B and Enterprise has no limit. In those terms Qwen3 8B and 14B fit Basic, which includes a 30-day free trial, and Qwen3 32B and 30B-A3B fit Professional. Qwen2.5 72B is just over the 70B line, so it falls under Enterprise. Which Qwen models are in the catalog at any time is shown in the console, not in this post. Plans are on the pricing section.
Each catalog variant lists measured GPU memory, counted at the full context window, so the console figure will be higher than the 8K numbers here. The same sizing logic applies to other families: see Llama 70B VRAM requirements for the 70B class, and DeepSeek R1 hardware requirements for reasoning models. Google's Gemma family uses a sliding-window design that keeps the cache much smaller; see Gemma VRAM requirements.
For the platform itself, see the self-hosted LLM platform overview.
Figures are estimates from public file sizes and configs. Measure on your hardware before buying GPUs.