Skip to content

Qwen3 30B A3B on 2× RTX 3090: VRAM, speed and fit

Qwen3 30B A3B at 8-bit runs on the 2× RTX 3090 with 40K of context at about 129 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on 2× RTX 3090 · 8-bit (Q8_0) · Ollama / LM Studio · Chat

GPU memory
35.4 GB
of 48.0 GB
Context
40K
auto
Speed
129 tok/s
72 when full
First reply
0.3 s
2K prompt

Where the memory goes

Model
30.2 GB
Context
3.8 GB
Overhead
1.4 GB
OS
0.0 GB
Free
12.6 GB
Memory
35.4 GB of 48.0 GB
Context
40K tokens · auto
Speed
129 tok/s · 72 when full
First reply
0.3 s for a 2K prompt
Running cost
$11/mo at home · $61/mo on a cloud 2× RTX A4500

Validation: Validated

What Qwen3 30B A3B on 2× RTX 3090 was checked against
CheckTargetResultStatus
Fits in GPU memory48.0 GB usable35.4 GB neededPass
Headroom5% free12.6 GB free (26%)Pass
Context for chat16K or more40KPass
Writing speed15 tok/s or more129 tok/sPass
First replyunder 3 s, 2K prompt0.3 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
Qwen3 30B A3B
Parameters
30.5B, 3B active
Precision
8-bit (Q8_0)
Context
40K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
2× RTX 3090
GPU can use
48.0 GB of 48.0 GB

Running cost

At home
$11/mo
500 W at the wall
Renting instead
$61/mo
2× RTX A4500 at $0.50/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 21, 2026.

How to run it

llama.cpp
llama-server -hf Qwen/Qwen3-30B-A3B-GGUF
vLLM
vllm serve Qwen/Qwen3-30B-A3B
MLX (Mac)
mlx_lm.chat --model mlx-community/Qwen3-30B-A3B-4bit

Qwen3 30B A3B on similar machines

All machines: Qwen3 30B A3B VRAM requirements. Everything that runs here: what can the 2× RTX 3090 run?

Other models that run on 2× RTX 3090

Questions

Can Qwen3 30B A3B run on 2× RTX 3090?

Qwen3 30B A3B at 8-bit runs on the 2× RTX 3090 with 40K of context at about 129 tokens a second.

Report VR-0RQIT15, calculated Nov 22, 2026 from published specs, not benchmarked.