Qwen 3 VL
The most powerful vision-language model in the Qwen3 family. Combines Qwen 3's reasoning with state-of-the-art visual understanding. 2B-235B sizes for any hardware.
Best for frontier-class vision-language with thinking
Sizes 2B · 4B · 8B · 30B · 32B · 235B
Context 128K
License Apache 2.0
Min VRAM (default size, Q4) 2 GB
Rec VRAM 8 GB
What is Qwen 3 VL?
Qwen 3 VL is Alibaba's open-weights language model with vision support. Released in 2025-10, it ships in 6 sizes (2B, 4B, 8B, 30B, 32B, 235B) and is licensed Apache 2.0. The most popular use case is frontier-class vision-language with thinking.
VRAM and hardware
The smallest 2B size needs at least 2 GB VRAM; the largest 235B needs around 2 GB.
| Size | Min VRAM (Q4_K_M) | Recommended VRAM | Tokens/sec on 3090 |
|---|---|---|---|
| 2B | 2 GB | 3 GB | ~31 tok/s |
| 4B | 3 GB | 5 GB | ~28 tok/s |
| 8B | 6 GB | 9 GB | ~24 tok/s |
| 30B | 18 GB | 27 GB | ~18 tok/s |
| 32B | 20 GB | 30 GB | ~17 tok/s |
| 235B | 130 GB | 195 GB | ~11 tok/s |
How to run Qwen 3 VL locally
Option 1: Ollama (simplest)
ollama pull qwen3-vl
ollama run qwen3-vl Option 2: Mullama (production)
mullama pull qwen3-vl
mullama run qwen3-vl Option 3: llama.cpp (CLI)
# Download a GGUF from Hugging Face (search "qwen3-vl gguf")
./llama-cli -m qwen3-vl.Q4_K_M.gguf -p "Hello, AI!" Option 4: Python with Mullama or llama-cpp-python
from mullama import Model, Context
model = Model.load("qwen3-vl.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256)) What you can build with Qwen 3 VL
- vision
- document understanding
- agentic
- multimodal
Hardware it fits on
- Apple Silicon (8GB): 2B in MLX
- Apple Silicon (16GB): 4B in MLX
- Apple Silicon (32GB): 8B in MLX or GGUF
- Apple Silicon (64GB+): 235B in MLX
- RTX 3090 (24GB): 8B at Q4_K_M
- RTX 4090 (24GB): same as 3090, ~30% faster
- RTX 5090 (32GB): 8B at Q8_0
- 2× RTX 6000 Ada (96GB total): 235B at Q4_K_M
Related models in the Qwen family
- Qwen 2.5 — 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B · strongest 7B-32B on consumer hardware
- Qwen 2.5-Coder — 0.5B, 1.5B, 3B, 7B, 14B, 32B · best open-weight code model in 2024-2025
- Qwen 3 — 0.6B, 1.7B, 4B, 8B, 14B, 30B, 32B, 235B · best all-round local LLM in 2025-2026
- Qwen 3-Coder — 30B, 480B · long-context code agent workflows
Other models you might consider
- Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
- Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
- Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
- Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
- Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
- DeepSeek R1 (DeepSeek) — 1.5B, reasoning, math, step-by-step problem solving