Qwen2.5-Coder 32B (Q3_K_M) · 10 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for Qwen2.5-Coder 32B (Q3_K_M)?
Qwen2.5-Coder 32B (Q3_K_M) on 10 GPUs, measured first-party: NVIDIA RTX PRO 6000 Blackwell Workstation Edition leads at 75.78 tok/s, L4 trails at 12.1 tok/s, and it peaked at 17GB of VRAM.
Benchmarked weights: bartowski/Qwen2.5-Coder-32B-Instruct-GGUF
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
75.78 tok/s on Qwen2.5-Coder 32B (Q3_K_M), the ceiling. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 75.78 tok/s on Qwen2.5-Coder 32B (Q3_K_M)
- 96GB, clears the Qwen2.5-Coder 32B (Q3_K_M) floor
- Rentable by the hour rather than bought
- 600W board rating
- Datacenter or workstation hardware, not a retail purchase
What GPU Do You Need for Qwen2.5-Coder 32B (Q3_K_M)?, 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.
Qwen2.5-Coder 32B (Q3_K_M). Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 75.78 | 3104.4 | 0.22 | 351.0 W |
| NVIDIA B300 | 74.28 | 1087.7 | 0.15 | 481.5 W |
| NVIDIA B200 | 60.24 | 2310.8 | 0.12 | 495.1 W |
| NVIDIA H200 | 58.87 | 2097.6 | 0.16 | 370.4 W |
| NVIDIA H100 80GB HBM3 | 58.05 | 2079.4 | 0.17 | 351.1 W |
| NVIDIA L40S | 41.06 | 2215.3 | 0.15 | 277.5 W |
| NVIDIA A100 80GB SXM4 | 31.99 | 925.4 | 0.13 | 249.0 W |
| NVIDIA A100 40GB SXM4 | 31.04 | 900.8 | 0.15 | 209.7 W |
| NVIDIA A10G | 18.78 | 776.9 | 0.14 | 133.3 W |
| NVIDIA L4 | 12.05 | 644.2 | 0.18 | 66.8 W |
What the numbers show. Across 10 GPUs measured on our own bench, RTX PRO 6000 Blackwell Workstation Edition is fastest at 75.8 tok/s. The slowest, L4, manages 12.1, so the spread is 6.3x from top to bottom.
How it compares. H100 80GB HBM3: Qwen2.5-Coder 32B (Q3_K_M) 58.05 tok/s, DeepSeek-R1-Distill-Qwen-32B-abliterated 73.48, Qwen2.5-32B 73.5 (33B), Qwen3-32B 74.07 (33B), Hermes-4-70B 41.2. 3 of 4 beat Qwen2.5-Coder 32B (Q3_K_M) here.
Cost on a rented GPU. 1M generated tokens of Qwen2.5-Coder 32B (Q3_K_M): $3.94 on a RTX PRO 6000 Blackwell Workstation Edition ($1.08/hr, 3.7 hours).
Qwen2.5-Coder 32B (Q3_K_M): cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 75.78 | $3.94 |
| NVIDIA A100 40GB SXM4 | $0.47/hr | 31.04 | $4.22 |
| NVIDIA L40S | $0.79/hr | 41.06 | $5.34 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 31.99 | $8.22 |
| NVIDIA L4 | $0.44/hr | 12.05 | $10.14 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 58.05 | $10.22 |
| NVIDIA H200 | $3.59/hr | 58.87 | $16.94 |
| NVIDIA B300 | $6.94/hr | 74.28 | $25.95 |
| NVIDIA B200 | $5.98/hr | 60.24 | $27.57 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Qwen2.5-Coder 32B (Q3_K_M). 30+ tok/s: 8 (RTX PRO 6000 Blackwell Workstation Edition, B300, B200); 10-30 tok/s: 2 (A10G, L4). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Qwen2.5-Coder 32B (Q3_K_M) writes anything it reads the input: 3104.4 tok/s on the RTX PRO 6000 Blackwell Workstation Edition (1.3s for a 4,000-token prompt), 644.2 on the L4 (6.2s). Long documents and big code files feel this number more than the generation speed.
VRAM for Qwen2.5-Coder 32B (Q3_K_M). Measured peak 15.7GB, so 24GB is the smallest common card size; smallest card it ran on: A10G (24GB).
Power on Qwen2.5-Coder 32B (Q3_K_M). Most efficient: RTX PRO 6000 Blackwell Workstation Edition, 351W, 1.29 kWh per 1M generated tokens. Hungriest: B200, 495W, 2.28 kWh. At $0.15/kWh: $0.19 per 1M generated tokens.
Our verdict
Fastest on Qwen2.5-Coder 32B (Q3_K_M): NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 75.78 tok/s. Cheapest to rent per job: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, $3.94 per 1M generated tokens.
FAQ
- What GPU do I need to run Qwen2.5-Coder 32B (Q3_K_M)?
- About 16GB. Smallest card that ran it: NVIDIA A10G (24GB).
- How much does it cost to run Qwen2.5-Coder 32B (Q3_K_M) in the cloud?
- $3.94 per 1M generated tokens on a NVIDIA RTX PRO 6000 Blackwell Workstation Edition at $1.08/hr, cheapest of 9 rentable cards we measured.
- Can I run Qwen2.5-Coder 32B (Q3_K_M) on a 12GB, 16GB or 24GB card?
- It used 15.7GB at the precision we tested. 12GB: no; 16GB: no; 24GB: yes.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for Qwen2.5-Coder 32B (Q3_K_M)?
- The H100 80GB HBM3: 58.05 vs 31.99 tok/s, 81% faster on our bench.
How we test
llama.cpp llama-bench at Q3_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.