Qwen3-VL 235B-A22B has 235.7B parameters (22B active per token). In BF16 its weights alone take 439 GiB; quantized to 4-bit, about 132 GiB. Each 32K-token request adds 5.88 GiB of KV cache. Cheapest way to run it today: 7 × GeForce RTX 3090 (24 GB) at $0.756 per hour. It reads text, images and video and calls tools.
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
How many GPUs to run Qwen3-VL 235B-A22B
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
BF16 / FP16
8-bit
4-bit
GeForce RTX 4090 24 GB
24 (several machines)
16 (several machines)
7
GeForce RTX 5090 32 GB
16 (several machines)
8
5
RTX PRO 6000 Blackwell 96 GB
5
3
2
Mac M5 Max 128 GB 96 GB usable
5
3
2
Ryzen AI Max+ 395 128 GB 96 GB usable
5
3
2
DGX Spark (GB10) 128 GB
4
2
2
NVIDIA H100 80 GB
7
4
2
NVIDIA H200 141 GB
4
2
2
NVIDIA B200 (HGX) 180 GB
3
2
1
8-bit: FP8 with vLLM (INT8 on GPUs without FP8), Q8_0 with llama.cpp. 4-bit: AWQ / GPTQ with vLLM, Q4_K_M with llama.cpp, both counted at the Q4_K_M size.
Cheapest way to run Qwen3-VL 235B-A22B today
For each GPU: the fewest cards that fit Qwen3-VL 235B-A22B with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-11-06).
On-demand prices from AWS, Google Cloud, Azure, RunPod, Vast.ai, Oracle Cloud, OVHcloud, Scaleway, Verda, Akash and UpCloud, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
vLLM and SGLang apply these defaults unless a request sets its own. With llama.cpp, pass them yourself: --temp 0.7 --top-p 0.8 --top-k 20
KV cache: how memory grows with context
Architecture: 94 layers · 94 full attention (4 KV × 128). Every layer keeps keys and values for every token (grouped-query attention), so the cache grows linearly with context. Only 22B parameters are active per token, but all 236B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
1.47 GiB
0.73 GiB
32K tokens
5.88 GiB
2.94 GiB
128K tokens
23.5 GiB
11.8 GiB
256K tokens
47.0 GiB
23.5 GiB
Qwen3-VL 235B-A22B VRAM FAQ
How much VRAM does Qwen3-VL 235B-A22B need?
In BF16 the weights alone take 439 GiB (471 GB). In FP8 that is 221 GiB, and about 132 GiB with 4-bit quantization (Q4_K_M). Each request then adds KV cache: 1.47 GiB at 8K tokens and 5.88 GiB at 32K (BF16 cache).
Can Qwen3-VL 235B-A22B run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 7 RTX 4090s. In 4-bit it fits on a single NVIDIA B200 (HGX) (180 GB).
Can Qwen3-VL 235B-A22B run on an RTX 5090 (32 GB)?
No. Even in 4-bit it needs 5 RTX 5090s.
Can Qwen3-VL 235B-A22B run on a Mac?
Yes. In 4-bit (Q4_K_M with llama.cpp) it needs a Mac with 192 GB of unified memory or more, of which macOS gives the GPU about 144 GB, for one request with an 8K-token context. In 8-bit (Q8_0) it needs 512 GB.
Can Qwen3-VL 235B-A22B run on a DGX Spark or a Ryzen AI Max+ 395 (128 GB)?
No. Even in 4-bit it needs 2 DGX Spark units or 2 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
What is the cheapest way to run Qwen3-VL 235B-A22B?
On 2026-11-06, the cheapest on-demand setup is 7 × GeForce RTX 3090 (24 GB) with Q4_K_M weights on Vast.ai, at $0.756 per hour (about $551.88 per month). Next: 3 × RTX A6000 (48 GB) with Q4_K_M weights on Vast.ai, at $0.966 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does Qwen3-VL 235B-A22B need?
With one request and an 8K-token context: 7 in BF16, 4 in FP8 and 2 in 4-bit (vLLM, 90% of memory usable). Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Can Qwen3-VL 235B-A22B read images?
Yes. Qwen3-VL 235B-A22B reads text, images and video. Images add tokens to the context, so they also add KV cache.
Does Qwen3-VL 235B-A22B support tool calling?
Yes, Qwen3-VL 235B-A22B supports tool (function) calling: it can ask your application to run a function and use the result.
What temperature should I use with Qwen3-VL 235B-A22B?
The official defaults are temperature 0.7, top-p 0.8, top-k 20.
What is Qwen3-VL 235B-A22B's knowledge cutoff?
March 2025, according to OpenRouter: it knows little about later events.
Qwen3-VL 235B-A22B VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/qwen3-vl-235b-a22b.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.
Alibaba, China · Released September 22, 2025