Qwen3 30B A3B Instruct 2507 · 8 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for Qwen3 30B A3B Instruct 2507?
Qwen3 30B A3B Instruct 2507 on 8 GPUs, measured first-party: NVIDIA GeForce RTX 5090 leads at 352.3 tok/s, L4 trails at 99.7 tok/s, and it peaked at 19GB of VRAM.
Benchmarked weights: unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF
NVIDIA GeForce RTX 5090
352.3 tok/s on Qwen3 30B A3B Instruct 2507, the ceiling. Measured on our bench. 32GB of VRAM, $1,999 at launch.
- 352.3 tok/s on Qwen3 30B A3B Instruct 2507
- 32GB, clears the Qwen3 30B A3B Instruct 2507 floor
- 575W board rating
NVIDIA GeForce RTX 4090
263.2 tok/s on Qwen3 30B A3B Instruct 2507, fastest card you can buy at retail. Measured on our bench. 24GB of VRAM, $1,599 at launch.
- 263.2 tok/s on Qwen3 30B A3B Instruct 2507
- 24GB, clears the Qwen3 30B A3B Instruct 2507 floor
- 450W board rating
NVIDIA GeForce RTX 3090
213.5 tok/s on Qwen3 30B A3B Instruct 2507, lowest launch price that still fits. Measured on our bench. 24GB of VRAM, $1,499 at launch.
- 213.5 tok/s on Qwen3 30B A3B Instruct 2507
- 24GB, clears the Qwen3 30B A3B Instruct 2507 floor
- 350W board rating
NVIDIA GeForce RTX 3090 Ti
230.3 tok/s on Qwen3 30B A3B Instruct 2507, most speed per dollar. Measured on our bench. 24GB of VRAM, $1,999 at launch. That is 115.2 tok/s per $1,000 of launch price.
- 230.3 tok/s on Qwen3 30B A3B Instruct 2507
- 24GB, clears the Qwen3 30B A3B Instruct 2507 floor
- 450W board rating
What GPU Do You Need for Qwen3 30B A3B Instruct 2507?, tok/s by GPU
Efficiency: tok/s per 100W drawn
Power is the average pulled during the run, sampled at 1Hz. The fastest card is often not the one here, and for anything left running this is the number that shows up on the bill.
Value: tok/s per $1,000 of MSRP
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
Qwen3 30B A3B Instruct 2507. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 | 352.3 | 11458.3 | 1.87 | 188.8 W |
| NVIDIA H100 80GB HBM3 | 299.8 | 7192 | 1.81 | 165.5 W |
| NVIDIA GeForce RTX 4090 | 263.2 | 9194.6 | 2.06 | 127.6 W |
| NVIDIA GeForce RTX 3090 Ti | 230.3 | 5228.8 | 1.22 | 189.5 W |
| NVIDIA L40S | 222.5 | 7767.2 | 1.64 | 135.4 W |
| NVIDIA GeForce RTX 3090 | 213.5 | 4345.5 | 1.17 | 181.7 W |
| NVIDIA A10G | 147.3 | 2882.7 | 1.6 | 92.0 W |
| NVIDIA L4 | 99.74 | 2440.4 | 2.22 | 45.0 W |
What the numbers show. Across 8 GPUs measured on our own bench, RTX 5090 is fastest at 352 tok/s. The slowest, L4, manages 99.7, so the spread is 3.5x from top to bottom. L4 is the most efficient, 99.7 tok/s at 45W.
About Qwen3 30B A3B Instruct 2507. Qwen3 30B A3B Instruct 2507: from Qwen, 31B parameters, on Hugging Face since July 2025, Apache 2.0 licence. 1,941,973 downloads in the last 30 days and 3 community quantizations.
How it compares. H100 80GB HBM3: Qwen3 30B A3B Instruct 2507 299.8 tok/s, Qwen3 30B A3B 283.8 (31B), Qwen3-Coder 30B A3B 296.2 (31B), GLM-4.7-Flash 186.1 (31B), Gemma 4 31B 76.59 (31B). Qwen3 30B A3B Instruct 2507 beats all 4 here.
Cost on a rented GPU. 1M generated tokens of Qwen3 30B A3B Instruct 2507: $0.16 on a RTX 3090 ($0.12/hr, 78 min), $0.31 on a RTX 5090 ($0.39/hr, 47 min, 1.9x the cost).
Qwen3 30B A3B Instruct 2507: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA GeForce RTX 3090 | $0.12/hr | 213.5 | $0.16 |
| NVIDIA GeForce RTX 5090 | $0.39/hr | 352.3 | $0.31 |
| NVIDIA GeForce RTX 3090 Ti | $0.27/hr | 230.3 | $0.33 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 263.2 | $0.35 |
| NVIDIA L40S | $0.79/hr | 222.5 | $0.99 |
| NVIDIA L4 | $0.44/hr | 99.74 | $1.23 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 299.8 | $1.98 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Qwen3 30B A3B Instruct 2507. 30+ tok/s: 8 (RTX 5090, RTX 4090, RTX 3090 Ti). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Qwen3 30B A3B Instruct 2507 writes anything it reads the input: 11458.3 tok/s on the RTX 5090 (0.3s for a 4,000-token prompt), 9194.6 on the RTX 4090 (0.4s), 2440.4 on the L4 (1.6s). Long documents and big code files feel this number more than the generation speed.
VRAM for Qwen3 30B A3B Instruct 2507. 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 24GB (tested), Q2_K 15GB, Q3_K_M 20GB, Q5_K_M 29GB, Q6_K 33GB.
Power on Qwen3 30B A3B Instruct 2507. Most efficient: L4, 45W, 0.13 kWh per 1M generated tokens. Hungriest: RTX 3090 Ti, 190W, 0.23 kWh. At $0.15/kWh: $0.019 per 1M generated tokens.
Our verdict
Fastest on Qwen3 30B A3B Instruct 2507: NVIDIA GeForce RTX 5090, 352.3 tok/s. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3090 ($1,499, 213.5 tok/s). Cheapest to rent per job: NVIDIA GeForce RTX 3090, $0.16 per 1M generated tokens.
FAQ
- What GPU do I need to run Qwen3 30B A3B Instruct 2507?
- About 19GB. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3090 (24GB, 213.5 tok/s).
- How fast is Qwen3 30B A3B Instruct 2507 on the NVIDIA GeForce RTX 5090?
- 352.3 tok/s at 189W, faster than every datacenter card we ran it on.
- How much does it cost to run Qwen3 30B A3B Instruct 2507 in the cloud?
- $0.16 per 1M generated tokens on a NVIDIA GeForce RTX 3090 at $0.12/hr, cheapest of 7 rentable cards we measured.
- Can I run Qwen3 30B A3B Instruct 2507 on a 12GB, 16GB or 24GB card?
- It used 17.8GB at the precision we tested. 12GB: no; 16GB: only at Q2_K (15GB); 24GB: yes.
- Is the RTX 4090 or the RTX 3090 faster for Qwen3 30B A3B Instruct 2507?
- The RTX 4090: 263.2 vs 213.5 tok/s, 23% faster on our bench.
- Should I buy or rent a RTX 5090 for Qwen3 30B A3B Instruct 2507?
- Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 6,518x 1M generated tokens of Qwen3 30B A3B Instruct 2507. Buy only if you'll run more than that.
How we test
llama.cpp llama-bench at Q4_K_M, 512-token prompt and 128 generated tokens, three runs after a warmup, full GPU offload, with power and VRAM sampled throughout. Token generation is memory-bandwidth-bound, so the ranking tracks bandwidth closely, which makes it a fair guide to cards we haven't run yet.