Best GPU for Qwen3 32B: VRAM Requirements
Qwen3 32B needs about 74.5 GB of GPU memory in BF16 and 20.4 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
Qwen3 32B needs about 74.5 GB of VRAM in FP16, 38.4 GB in FP8 and 20.4 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on one NVIDIA RTX PRO 6000 Blackwell 96GB; a 4-bit build fits on one NVIDIA GeForce RTX 4090 24GB.
Qwen3 32B is a 32.8B-parameter language model from Alibaba Cloud (Qwen team), released in April 2025, with a 32,768-token native context window that extends to 131,072 tokens with YaRN and published under the Apache 2.0. Its weights alone take 65.6 GB in BF16 and 16.4 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.
Recommended GPUs for Qwen3 32B
Best for inference
1 GPU × NVIDIA L40S 48GB
FP8 or INT8 weights, 38.4 GB at 8,192 tokens.
Best for fine-tuning
1 GPU × NVIDIA H200 141GB
LoRA on BF16 weights, 85.3 GB.
Budget pick
1 GPU × NVIDIA GeForce RTX 4090 24GB
Smallest card class that holds a 4-bit build, 20.4 GB.
Qwen3 32B VRAM requirements by precision
One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.
| Precision | Weights | KV cache | Total for inference |
|---|---|---|---|
| BF16 / FP16 | 65.6 GB | 2.1 GB | 74.5 GB |
| FP8 | 32.8 GB | 2.1 GB | 38.4 GB |
| INT4 | 16.4 GB | 2.1 GB | 20.4 GB |
Qwen3 32B VRAM by task
Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.
| Task | GPU memory |
|---|---|
| Inference at the full 131,072-token context (FP8) | 73.9 GB |
| QLoRA (4-bit base) | 21.3 GB |
| LoRA (BF16 base) | 85.3 GB |
| Full fine-tuning (BF16, AdamW) | 577.3 GB |
GPUs that fit Qwen3 32B
Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.
| GPU | Memory | BF16 inference (74.5 GB) | FP8 inference (38.4 GB) | INT4 inference (20.4 GB) | QLoRA (21.3 GB) | LoRA (85.3 GB) |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5080 16GB | 16 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA GeForce RTX 4090 24GB | 24 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA L4 24GB | 24 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA GeForce RTX 5090 32GB | 32 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA Tesla V100 32GB | 32 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA A100 PCIe 40GB | 40 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA L40S 48GB | 48 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA Quadro RTX 8000 48GB | 48 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA RTX 6000 Ada 48GB | 48 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA RTX A6000 48GB | 48 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA A100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA A100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA H100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA H100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H200 141GB | 141 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B200 180GB | 180 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B300 288GB | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA GB300 NVL | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
What is Qwen3 32B?
Qwen3 32B is the largest dense model of the Qwen3 family, with the same thinking and non-thinking modes as its smaller siblings. Its Apache 2.0 license and size make it a common single-server choice for reasoning and agent workloads.
- Developer
- Alibaba Cloud (Qwen team)
- Released
- April 2025
- Parameters
- 32.8B
- Context window
- 32,768 tokens native, 131,072 with YaRN
- License
- Apache 2.0
The model card supports 131,072 tokens with YaRN scaling; the native window is 32,768 tokens.
Source: Qwen3 32B model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does Qwen3 32B need?
- For inference with an 8,192-token context and one request, Qwen3 32B needs about 74.5 GB in BF16, 38.4 GB in FP8, 20.4 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can Qwen3 32B run on a single GPU?
- Yes. In BF16 it needs about 74.5 GB, which fits one NVIDIA RTX PRO 6000 Blackwell 96GB (96 GB).
- How much GPU memory does fine-tuning Qwen3 32B take?
- QLoRA (4-bit base): about 21.3 GB; LoRA (BF16 base): about 85.3 GB; Full fine-tuning (BF16, AdamW): about 577.3 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for Qwen3 32B?
- Best for inference: 1 GPU × NVIDIA L40S 48GB (FP8 or INT8 weights, 38.4 GB at 8,192 tokens). Best for fine-tuning: 1 GPU × NVIDIA H200 141GB (LoRA on BF16 weights, 85.3 GB). Budget pick: 1 GPU × NVIDIA GeForce RTX 4090 24GB (Smallest card class that holds a 4-bit build, 20.4 GB). Check live offers before booking, as availability changes.
- Can Qwen3 32B run on an RTX 4090?
- Yes. In 4-bit it needs about 20.4 GB, which fits one RTX 4090 (24 GB). FP16 needs about 74.5 GB, 4 of these cards.
- Can Qwen3 32B run on an RTX 5090?
- Yes. In 4-bit it needs about 20.4 GB, which fits one RTX 5090 (32 GB). FP16 needs about 74.5 GB, 4 of these cards.
Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.
Explore the live market
Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.