AlibabaQwen 3Qwen3VLMoeForConditionalGeneration

Qwen3 VL 30B A3B Instruct — Hardware Requirements & GPU Compatibility

VisionFunctions

Qwen3 VL 30B A3B Instruct is Alibaba's 31-billion-parameter mixture-of-experts vision-language model in the Qwen 3 lineup, with about 3 billion parameters active per token (the A3B in its name). Because only the active experts run for each token, inference is faster than a similarly sized dense model, while all the weights still need to fit in memory. It handles images and video alongside text, supporting visual question answering, document and chart reading, and multi-image reasoning. At this size, local inference calls for quantization and a single high-end consumer GPU. The model supports a 262K token context window, suited to long documents, video transcripts, or extended visual conversations. It is released under the Apache 2.0 license, allowing unrestricted commercial and research use, and was published in September 2025 as part of Qwen's third-generation vision-language family.

31.1B params · MoE3.9B active262K contextApache 2.0Sep 2025444.5K downloads627 likes435.4K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 12 GB (IQ2_M · 16K)Best on 24 GB (Q4_K_M · ~159 tok/s · 64K)Macs from 16 GB (IQ2_M · ~52 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Qwen3 VL 30B A3B Instruct

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX A5000 (24 GB) (4 GB headroom) · Q4_K_M

Generation speed: ~135 tok/s; generation speed
Cost per 1M output tokens: $0.33; per 1M output tokens

Deployon RunPodBest valuefrom $0.16/hr

$5+ bonus for new users

Compare GPUs →

Run Qwen3 VL 30B A3B Instruct

19.7 GB VRAM at 16K context · fits the RTX 4090 (4.3 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run qwen3-vl:30b-a3b-instruct
Official Ollama library build. For the exact Q4_K_M file from Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF: ollama run hf.co/Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and search qwen/qwen3-vl-30b, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get qwen/qwen3-vl-30b@q4_k_m
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve Qwen/Qwen3-VL-30B-A3B-Instruct --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~68 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model lmstudio-community/Qwen3-VL-30B-A3B-Instruct-MLX-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model lmstudio-community/Qwen3-VL-30B-A3B-Instruct-MLX-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$Qwen/Qwen3-VL-30B-A3B-Instruct-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX A5000 (24 GB)CheapestCheapest4 GB free	~135 tok/s	$0.33	Deploy on RunPodfrom $0.16/hr (opens in new tab)	24 GB · 4 GB free · ~135 tok/s · $0.33 / 1M tokens
RTX 3090 (24 GB)4 GB free	~153 tok/s	$0.31	Deploy on Vast.aifrom $0.17/hr (opens in new tab)	24 GB · 4 GB free · ~153 tok/s · $0.31 / 1M tokens
TITAN RTX (24 GB)4 GB free	~124 tok/s	$0.48	Deploy on Vast.aifrom $0.22/hr (opens in new tab)	24 GB · 4 GB free · ~124 tok/s · $0.48 / 1M tokens
RTX A5000 (24 GB)4 GB free	~135 tok/s	$0.47	Deploy on Vast.aifrom $0.23/hr (opens in new tab)	24 GB · 4 GB free · ~135 tok/s · $0.47 / 1M tokens
RTX 3090 Ti (24 GB)4 GB free	~159 tok/s	$0.47	Deploy on Vast.aifrom $0.27/hr (opens in new tab)	24 GB · 4 GB free · ~159 tok/s · $0.47 / 1M tokens
RTX PRO 4500 (32 GB)12 GB free	~149 tok/s	$0.64	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 12 GB free · ~149 tok/s · $0.64 / 1M tokens
RTX 4090 (24 GB)4 GB free	~159 tok/s	$0.59	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 4 GB free · ~159 tok/s · $0.59 / 1M tokens
A40 (48 GB)28 GB free	~126 tok/s	$0.77	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 28 GB free · ~126 tok/s · $0.77 / 1M tokens

Compare GPUs →

How Much VRAM Does Qwen3 VL 30B A3B Instruct Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
8.9 GB	10.5 GB	15.3 GB	10.3 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 256K context · ~230 tok/s
9.7 GB	11.2 GB	16.1 GB	8.7 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 256K context · ~222 tok/s
10.1 GB	11.6 GB	16.5 GB	8.7 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 256K context · ~219 tok/s
10.9 GB	12.4 GB	17.2 GB	10.8 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 256K context · ~212 tok/s
12.4 GB	14 GB	18.8 GB	12.9 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 128K context · ~199 tok/s
12.8 GB	14.3 GB	19.2 GB	10.5 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 128K context · ~196 tok/s
13.2 GB	14.7 GB	19.6 GB	12.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 128K context · ~194 tok/s
13.6 GB	15.1 GB	20 GB	13.3 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 128K context · ~191 tok/s
13.6 GB	15.1 GB	20 GB	11.3 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 128K context · ~191 tok/s
14.0 GB	15.5 GB	20.3 GB	13.3 GB	3-bit small — 3-bit small quantization	Fits: 128K context · ~188 tok/s
14.4 GB	15.9 GB	20.7 GB	14.1 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 128K context · ~186 tok/s
15.6 GB	17.1 GB	21.9 GB	14.7 GB	3-bit medium — 3-bit medium quantization	Fits: 128K context · ~178 tok/s
15.9 GB	17.5 GB	22.3 GB	17.4 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 128K context · ~176 tok/s
16.3 GB	17.8 GB	22.7 GB	14.6 GB	3-bit large — 3-bit large quantization	Fits: 128K context · ~174 tok/s
17.1 GB	18.6 GB	23.4 GB	16.4 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 128K context · ~169 tok/s
17.9 GB	19.4 GB	24.2 GB	19.2 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 64K context · ~165 tok/s
17.9 GB	19.4 GB	24.2 GB	17.5 GB	4-bit small — 4-bit small quantization	Fits: 64K context · ~165 tok/s
17.9 GB	19.4 GB	24.2 GB	17.3 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 64K context · ~165 tok/s
19.0 GB	20.6 GB	25.4 GB	18.6 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 64K context · ~159 tok/s
19.4 GB	20.9 GB	25.8 GB	18.9 GB	4-bit large — 4-bit large quantization	Fits: 64K context · ~157 tok/s
21.8 GB	23.3 GB	28.1 GB	21.1 GB	5-bit small — 5-bit small quantization	Fits: 32K context · ~147 tok/s
22.5 GB	24.1 GB	28.9 GB	21.7 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 16K context · ~144 tok/s
22.9 GB	24.4 GB	29.3 GB	21.9 GB	5-bit large — 5-bit large quantization	Fits: 16K context · ~142 tok/s
26.0 GB	27.5 GB	32.4 GB	25.1 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 26.7 GB
31.5 GB	33 GB	37.8 GB	32.5 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 32.2 GB
62.5 GB	64.1 GB	68.9 GB	61.1 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 63.2 GB

Selected Q4_K_M

4-bit medium quantization

19.7GBat 16K context

Weights: 18.6 GB
KV cache: 0.8 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 4.3 GB free · ~159 tok/s

≥50 tok/s needs ≥211 GB/s · RTX 5090 ~213, RTX 3090 Ti ~159

Which Hardware Can Run Qwen3 VL 30B A3B Instruct?

Q4_K_M · 19.0 GB

Qwen3 VL 30B A3B Instruct (Q4_K_M) requires 19.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 25+ GB is recommended. Using the full 262K context window can add up to 12.8 GB, bringing total usage to 31.8 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 122 devices with unified memory can run Qwen3 VL 30B A3B Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M1 Max (32 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060Doesn't fit — smallest quant needs 9.7 GB8 GB tier →

12 GB · RTX 3060 12GBIQ2_M · 11.6 GB · ~118 tok/s · 16K11.6~11816K12 GB tier →

16 GB · RTX 5060 Ti 16GBIQ3_M · 15.1 GB · ~113 tok/s · 32K15.1~11332K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 19.7 GB · ~159 tok/s · 64K19.7~15964K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 19.7 GB · ~26 tok/s · 128K19.7~26128KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 19.7 GB · ~63 tok/s · 256K19.7~63256KMacs →

Alternatives to Qwen3 VL 30B A3B Instruct

Qwen 3 family · 79 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Nanbeige4.3 33BNanbeige	—	23.7 GB	~32	same size
Hy Memory 30B A3BTencent	—	19.1 GB	~184	same size
Granite Switch 4.1 30BIBM	—	22.7 GB	~35	same size
Llm Jp 4.1 32B A3b Thinkingllm-jp	—	20.1 GB	~188	same size
Llm Jp 4.1 33B Thinkingllm-jp	—	22.5 GB	~35	same size
Agnes 3.0 FlashAgnes-AI	—	22.1 GB	~35	same size

Specifications

Publisher: Alibaba
Family: Qwen 3
Parameters: 31.1B
Active parameters: 3.9B per token
Architecture: Qwen3VLMoeForConditionalGeneration
Context Length: 262,144 tokens
Vocabulary Size: 151,936
Release Date: 2025-09-30
License: Apache 2.0

Resources

Frequently Asked Questions

How much VRAM does Qwen3 VL 30B A3B Instruct need?

Qwen3 VL 30B A3B Instruct requires 19.0 GB of VRAM at Q4_K_M, or 62.5 GB at BF16. Full 262K context adds up to 12.8 GB (31.8 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 31.1B × 4.8 bits ÷ 8 = 18.6 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 13.2 GB (at full 262K context)

VRAM usage by quantization

Q4_K_M

19.0 GB

Q4_K_M + full context

31.8 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Qwen3 VL 30B A3B Instruct?

Yes, at Q5_K_L (22.9 GB) or lower. Higher quantizations like Q6_K (26.0 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Qwen3 VL 30B A3B Instruct?

For Qwen3 VL 30B A3B Instruct, Q4_K_M (19.0 GB) offers the best balance of quality and VRAM usage. Q4_K_L (19.4 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 8.9 GB.

VRAM requirement by quantization

IQ2_XXS

8.9 GB~53%

IQ3_XS

13.2 GB~73%

Q3_K_L

16.3 GB~86%

Q4_K_M ★

19.0 GB~89%

Q4_K_L

19.4 GB~90%

BF16

62.5 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Qwen3 VL 30B A3B Instruct on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Qwen3 VL 30B A3B Instruct, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 19.0 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M5 Pro (24 GB).

Can I run Qwen3 VL 30B A3B Instruct locally?

Yes — Qwen3 VL 30B A3B Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Qwen3 VL 30B A3B Instruct?

At Q4_K_M, Qwen3 VL 30B A3B Instruct can reach ~99 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~159 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 2.4 GB active ÷ (8000 × 0.65) = 0.46 ms, plus 48 layers × 0.055 ms = 2.64 ms, so 1000 ÷ 3.10 ms = ~323 tok/s

Estimated speed at Q4_K_M (19.0 GB)

NVIDIA B200

~323 tok/s

NVIDIA GeForce RTX 4090

~159 tok/s

NVIDIA B300

~323 tok/s

NVIDIA GH200 Grace Hopper Superchip

~295 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Qwen3 VL 30B A3B Instruct?

At Q4_K_M, the download is about 18.64 GB. The full-precision BF16 version is 62.14 GB. The smallest option (IQ2_XXS) is 8.54 GB.

Which GPUs can run Qwen3 VL 30B A3B Instruct?

8 consumer GPUs can run Qwen3 VL 30B A3B Instruct at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.

Which devices can run Qwen3 VL 30B A3B Instruct?

122 devices with unified memory can run Qwen3 VL 30B A3B Instruct at Q4_K_M (19.0 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.