Best GPU for Qwen3 235B A22B: VRAM Requirements
Qwen3 235B A22B needs about 518.7 GB of GPU memory in BF16 and 131 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
Qwen3 235B A22B needs about 518.7 GB of VRAM in FP16, 260.2 GB in FP8 and 131 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on 4 × NVIDIA B200 180GB; a 4-bit build fits on one NVIDIA B200 180GB.
Qwen3 235B A22B is a mixture-of-experts language model with 235B total and 22B active parameters per token 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 470 GB in BF16 and 117.5 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 235B A22B
Best for inference
2 GPUs × NVIDIA B200 180GB
FP8 or INT8 weights, 260.2 GB at 8,192 tokens.
Best for fine-tuning
4 GPUs × NVIDIA B200 180GB
LoRA on BF16 weights, 611 GB.
Budget pick
2 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB
Smallest card class that holds a 4-bit build, 131 GB.
Qwen3 235B A22B 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 | 470 GB | 1.6 GB | 518.7 GB |
| FP8 | 235 GB | 1.6 GB | 260.2 GB |
| INT4 | 117.5 GB | 1.6 GB | 131 GB |
Qwen3 235B A22B 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) | 286.3 GB |
| QLoRA (4-bit base) | 152.8 GB |
| LoRA (BF16 base) | 611 GB |
| Full fine-tuning (BF16, AdamW) | 4,136 GB |
GPUs that fit Qwen3 235B A22B
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 (518.7 GB) | FP8 inference (260.2 GB) | INT4 inference (131 GB) | QLoRA (152.8 GB) | LoRA (611 GB) |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 4090 24GB | 24 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA L4 24GB | 24 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA GeForce RTX 5090 32GB | 32 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA Tesla V100 32GB | 32 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA A100 PCIe 40GB | 40 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA L40S 48GB | 48 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA Quadro RTX 8000 48GB | 48 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA RTX 6000 Ada 48GB | 48 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA RTX A6000 48GB | 48 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA A100 PCIe 80GB | 80 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA A100 SXM 80GB | 80 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA H100 PCIe 80GB | 80 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA H100 SXM 80GB | 80 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | 8 GPUs | 4 GPUs | 2 GPUs | 2 GPUs | 8 GPUs |
| NVIDIA H200 141GB | 141 GB | 4 GPUs | 2 GPUs | 1 GPU | 2 GPUs | 8 GPUs |
| NVIDIA B200 180GB | 180 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA B300 288GB | 288 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA GB300 NVL | 288 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 4 GPUs |
What is Qwen3 235B A22B?
Qwen3 235B A22B is the flagship mixture-of-experts model of the Qwen3 family: 128 experts with 8 active per token, so it computes about 22 billion parameters per token while holding 235 billion in memory. It supports thinking and non-thinking modes in one checkpoint.
- Developer
- Alibaba Cloud (Qwen team)
- Released
- April 2025
- Parameters
- 235B total, 22B active per token
- 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 235B A22B model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does Qwen3 235B A22B need?
- For inference with an 8,192-token context and one request, Qwen3 235B A22B needs about 518.7 GB in BF16, 260.2 GB in FP8, 131 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can Qwen3 235B A22B run on a single GPU?
- Yes. In INT4 it needs about 131 GB, which fits one NVIDIA B200 180GB (180 GB). Higher precisions need a larger card or several GPUs.
- How much GPU memory does fine-tuning Qwen3 235B A22B take?
- QLoRA (4-bit base): about 152.8 GB; LoRA (BF16 base): about 611 GB; Full fine-tuning (BF16, AdamW): about 4,136 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for Qwen3 235B A22B?
- Best for inference: 2 GPUs × NVIDIA B200 180GB (FP8 or INT8 weights, 260.2 GB at 8,192 tokens). Best for fine-tuning: 4 GPUs × NVIDIA B200 180GB (LoRA on BF16 weights, 611 GB). Budget pick: 2 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB (Smallest card class that holds a 4-bit build, 131 GB). Check live offers before booking, as availability changes.
- Can Qwen3 235B A22B run on an RTX 4090?
- Not on one card: even a 4-bit build needs about 131 GB, so it takes 8 RTX 4090 cards (192 GB combined) with the model split across them.
- Can Qwen3 235B A22B run on an RTX 5090?
- Not on one card: even a 4-bit build needs about 131 GB, so it takes 8 RTX 5090 cards (256 GB combined) with the model split across them.
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.