What GPU do I need to run Qwen/Qwen-72B?
72.3B parameters, published in BF16. View on Hugging Face
Set up Qwen-72BQwen-72B is published by Qwen on Hugging Face, with 3,244,398 downloads and 361 likes to date. It's a QWenLMHeadModel model built for text-generation, published natively in BF16.
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| BF16 | 134.6 GB | 161.6 GB | RTX A5000 | 7 | $1.23/hr |
| FP8 (quantized) | 67.3 GB | 80.8 GB | RTX PRO 6000 | 1 | $1.38/hr |
| cheaper alt. | RTX 4070 Super | 7 | $0.847/hr | ||
| INT4 (quantized) | 33.7 GB | 40.4 GB | RTX A6000 | 1 | $0.363/hr |
| cheaper alt. | RTX A5000 | 2 | $0.352/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run Qwen-72B at its published (BF16) precision: 7× RTX A5000, at $0.176/hr per GPU ($1.23/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
Qwen-72B: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Multi-head attention: every layer caches keys and values for every head. Cached per token: 2,621,440 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 10.0 GB |
| 32K | 80.0 GB |
Computed from the layer, head and window counts in the model's own configuration (Qwen/Qwen-72B), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
Qwen-72B: common questions
Can Qwen-72B run on a single GPU?
No. At BF16 it needs 161.6 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest capable card in the live feed is a 24.0 GB RTX A5000, and it takes 7 of them.
How many GPUs do I need to run Qwen-72B?
7 at BF16. It needs 161.6 GB of VRAM and the cheapest capable live offer is a 24.0 GB RTX A5000, so 7 of them come to $1.23/hr in total.
Does quantizing Qwen-72B lower the GPU bill?
Yes. At BF16 the cheapest live fit is 7 RTX A5000 cards at $1.23/hr. At INT4 (quantized) it drops to one RTX A6000 at $0.363/hr, provided a quantized checkpoint exists for it.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Qwen 1 models
- Qwen-7B-Chat (7.7B, BF16)
- Qwen-7B (7.7B, BF16)
Alternatives at this size
Other models for text-generation within about a third of Qwen-72B's 72.3B parameters, from other model lines.
- Llama-3.3-70B-Instruct (70.6B, BF16)
- Kimi-Linear-48B-A3B-Instruct (49.1B, BF16)
- Hunyuan-A13B-Instruct (80.4B, BF16)
- Apertus-70B-Instruct-2509 (70.6B, BF16)
- Le_Triomphant-ECE-TW3 (72.3B, BF16)
More on Qwen-72B
Fits on a 192 GB GPU at BF16: every model that fits in 192 GB.
Fits on a 96 GB GPU at FP8: every model that fits in 96 GB.
Fits on a 48 GB GPU at INT4: every model that fits in 48 GB.
Best chat and assistants models: how Qwen-72B ranks against the rest.
Compare Qwen-72B with Llama-3.3-70B-Instruct, Llama-3.1-70B-Instruct, Phi-3.5-MoE-instruct, Meta-Llama-3-70B.
Related reading: RTX A5000 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
