Qwen3 30B A3B · 25 GPUs measured first-party · llama.cpp Q4_K_M · Updated October 2026

What GPU Do You Need for Qwen3 30B A3B?

Qwen3 30B-A3B is the model that changes the local-LLM math. It's a mixture-of-experts design, 30B parameters in memory, only ~3B active per token, so it needs the VRAM of a big model but generates at small-model speed: 309 tok/s on the RTX PRO 6000 Blackwell in our tests, faster than the dense 8B on the same silicon. Measured on 25 GPUs, ~20GB peak VRAM at Q4_K_M.

Benchmarked weights: bartowski/Qwen_Qwen3-30B-A3B-GGUF

Fastest we measured

NVIDIA GeForce RTX 5090

344.7 tok/s on Qwen3 30B A3B, the ceiling. Measured on our bench. 32GB of VRAM, $1,999 at launch.

Pros

• 344.7 tok/s on Qwen3 30B A3B
• 32GB, clears the Qwen3 30B A3B floor

Cons

• 575W board rating

Best consumer card

NVIDIA GeForce RTX 4090

259.6 tok/s on Qwen3 30B A3B, fastest card you can buy at retail. Measured on our bench. 24GB of VRAM, $1,599 at launch.

Pros

• 259.6 tok/s on Qwen3 30B A3B
• 24GB, clears the Qwen3 30B A3B floor

Cons

• 450W board rating

Cheapest card that runs it

NVIDIA GeForce RTX 3090

202.3 tok/s on Qwen3 30B A3B, lowest launch price that still fits. Measured on our bench. 24GB of VRAM, $1,499 at launch.

Pros

• 202.3 tok/s on Qwen3 30B A3B
• 24GB, clears the Qwen3 30B A3B floor

Cons

• 350W board rating

Best value

NVIDIA GeForce RTX 3090 Ti

226.2 tok/s on Qwen3 30B A3B, most speed per dollar. Measured on our bench. 24GB of VRAM, $1,999 at launch. That is 113.2 tok/s per $1,000 of launch price.

Pros

• 226.2 tok/s on Qwen3 30B A3B
• 24GB, clears the Qwen3 30B A3B floor

Cons

• 450W board rating

This is where it gets good. In my tiering of the Qwen lineup, this is the model where local LLM work stops being a compromise. The MoE trick means you pay for capability in VRAM instead of speed, 20GB in memory, but per-token compute like a 3B, and the result is a strong 30B-class model that generates at 300+ tok/s on the right card. If you own a 24GB card, this should probably be your default daily model. It's the reason I tell people a used 24GB card is the smartest local-AI purchase: this model is what it unlocks.

What the measurements show. The RTX 5090 tops the chart at 345 tok/s, ahead of the RTX PRO 6000 Blackwell at 309. And does it at 132W, 2.34 tok/W, because the active-parameter count is so small the card never works hard. The H200 follows at 292 tok/s. Even the A10G, a modest cloud card, holds 150 tok/s. Compare that to the dense Qwen3 32B, where the fastest result we have anywhere is 84 tok/s: same memory class, nearly 4× the generation speed. The ~20GB floor means 24GB consumer cards, a $999 RX 7900 XTX, a used RTX 3090, clear it with room for context.

About Qwen3 30B A3B. Qwen3 30B A3B: from Qwen, 31B parameters, on Hugging Face since April 2025, Apache 2.0 licence. 1,872,180 downloads in the last 30 days and 2 community quantizations.

How it compares. H100 80GB HBM3: Qwen3 30B A3B 283.8 tok/s, Qwen3 30B A3B Instruct 2507 299.8 (31B), Qwen3-Coder 30B A3B 296.2 (31B), GLM-4.7-Flash 186.1 (31B), Gemma 4 31B 76.59 (31B). 2 of 4 beat Qwen3 30B A3B here.

Cost on a rented GPU. 1M generated tokens of Qwen3 30B A3B: $0.17 on a RTX 3090 ($0.12/hr, 82 min), $0.31 on a RTX 5090 ($0.39/hr, 48 min, 1.9x the cost).

Speed tiers for Qwen3 30B A3B. 30+ tok/s: 25 (RTX 5090, RTX 4090, RTX 3090 Ti). 30 tok/s is roughly where replies outpace reading.

Reading your prompt. Before Qwen3 30B A3B writes anything it reads the input: 10736.9 tok/s on the RTX 5090 (0.4s for a 4,000-token prompt), 9207.6 on the RTX 4090 (0.4s), 2404.6 on the Quadro RTX 8000 (1.7s). Long documents and big code files feel this number more than the generation speed.

VRAM for Qwen3 30B A3B. Measured peak 17.8GB, so 24GB is the smallest common card size; smallest card it ran on: RTX 4090 (24GB). With long context: Q4_K_M 20GB (tested), Q2_K 12GB, Q3_K_M 16GB, Q5_K_M 24GB, Q6_K 27GB.

Power on Qwen3 30B A3B. Most efficient: RTX PRO 4000 Blackwell, 74W, 0.11 kWh per 1M generated tokens. Hungriest: B200, 269W, 0.28 kWh. At $0.15/kWh: $0.017 per 1M generated tokens.

Our verdict

Qwen3 30B-A3B: 309 tok/s peak at just 132W, ~20GB floor. The MoE architecture makes it generate like a small model while reasoning like a big one, in our data it's nearly 4× faster than the dense 32B in the same memory class. If you have a 24GB card, this is the model that justifies it.

FAQ

Why is a 30B model this fast?

It's a mixture-of-experts: all 30B parameters sit in VRAM, but only ~3B activate per token. You pay the memory cost of a big model and the compute cost of a small one. Which is why it hit 309 tok/s where the dense Qwen3 32B tops out at 84 in our database.

What GPU do I need for Qwen3 30B-A3B?

~20GB measured peak at Q4_K_M, so a 24GB card: RX 7900 XTX ($999 MSRP), used RTX 3090, RTX 4090. 16GB cards don't fit it at this quant.

Qwen3 30B-A3B or Qwen3 32B dense?

For most local use, the MoE. It's nearly 4× faster in our measurements (309 vs 84 tok/s at each chart's top) in the same VRAM class. The dense 32B only earns its slot when its per-token depth genuinely wins on your tasks.

How efficient is it?

The best big-model efficiency we've measured outside coder-30B: 2.34 tok/W on the RTX PRO 6000 (309 tok/s at 132W). The low active-parameter count keeps the silicon loafing.

Is this the best model for a used RTX 3090?

It's the strongest argument for one. The 3090's 24GB clears the ~20GB floor, and the MoE speed means the aging bandwidth still delivers a fast experience, a big-model brain on a 2020 card.