Olmo-3.1-32B-Think · 10 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for Olmo-3.1-32B-Think?
Olmo-3.1-32B-Think on 10 GPUs, measured first-party: NVIDIA B300 leads at 87.85 tok/s, L4 trails at 12.3 tok/s, and it peaked at 20GB of VRAM.
Benchmarked weights: bartowski/allenai_Olmo-3.1-32B-Think-GGUF
NVIDIA B300
87.85 tok/s on Olmo-3.1-32B-Think, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 87.85 tok/s on Olmo-3.1-32B-Think
- 288GB, clears the Olmo-3.1-32B-Think floor
- Rentable by the hour rather than bought
- 1400W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
65.71 tok/s on Olmo-3.1-32B-Think, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 65.71 tok/s on Olmo-3.1-32B-Think
- 96GB, clears the Olmo-3.1-32B-Think 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 Olmo-3.1-32B-Think?, 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.
Olmo-3.1-32B-Think. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA B300 | 87.85 | 1408.6 | 0.18 | 476.1 W |
| NVIDIA B200 | 79.75 | 2592.3 | 0.16 | 503.2 W |
| NVIDIA H200 | 78.46 | 2295.1 | 0.23 | 336.3 W |
| NVIDIA H100 80GB HBM3 | 75.53 | 2373.3 | 0.22 | 344.0 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 65.71 | 3414.1 | 0.2 | 321.4 W |
| NVIDIA A100 80GB SXM4 | 46.49 | 1182.3 | 0.18 | 260.9 W |
| NVIDIA A100 40GB SXM4 | 43.67 | 1160.6 | 0.21 | 212.0 W |
| NVIDIA L40S | 34.11 | 2387.6 | 0.14 | 248.3 W |
| NVIDIA A10G | 22.13 | 800.8 | 0.17 | 131.9 W |
| NVIDIA L4 | 12.27 | 703.3 | 0.18 | 66.7 W |
What the numbers show. Across 10 GPUs measured on our own bench, B300 is fastest at 87.8 tok/s. The slowest, L4, manages 12.3, so the spread is 7.2x from top to bottom. H200 is the most efficient, 78.5 tok/s at 336W. Per dollar of launch price, A10G gives the most (7.9 tok/s per $1,000).
How it compares. H100 80GB HBM3: Olmo-3.1-32B-Think 75.53 tok/s, DeepSeek-R1 Distill 32B 75.8 (33B), Qwen2.5-Coder 32B 75.81 (33B), QwQ 32B 75.82, Gemma 4 31B 76.59 (31B). All 4 beat Olmo-3.1-32B-Think here.
Cost on a rented GPU. 1M generated tokens of Olmo-3.1-32B-Think: $3.00 on a A100 40GB SXM4 ($0.47/hr, 6.4 hours), $21.94 on a B300 ($6.94/hr, 3.2 hours, 7.3x the cost).
Olmo-3.1-32B-Think: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA A100 40GB SXM4 | $0.47/hr | 43.67 | $3.00 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 65.71 | $4.55 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 46.49 | $5.66 |
| NVIDIA L40S | $0.79/hr | 34.11 | $6.43 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 75.53 | $7.86 |
| NVIDIA L4 | $0.44/hr | 12.27 | $9.96 |
| NVIDIA H200 | $3.59/hr | 78.46 | $12.71 |
| NVIDIA B200 | $5.98/hr | 79.75 | $20.83 |
| NVIDIA B300 | $6.94/hr | 87.85 | $21.94 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Olmo-3.1-32B-Think. 30+ tok/s: 8 (B300, B200, H200); 10-30 tok/s: 2 (A10G, L4). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Olmo-3.1-32B-Think writes anything it reads the input: 3414.1 tok/s on the RTX PRO 6000 Blackwell Workstation Edition (1.2s for a 4,000-token prompt), 703.3 on the L4 (5.7s). Long documents and big code files feel this number more than the generation speed.
VRAM for Olmo-3.1-32B-Think. Measured peak 18.9GB, so 24GB is the smallest common card size; smallest card it ran on: A10G (24GB). With long context: Q4_K_M 25GB (tested), Q2_K 16GB, Q3_K_M 21GB, Q5_K_M 30GB, Q6_K 34GB.
Power on Olmo-3.1-32B-Think. Most efficient: H200, 336W, 1.19 kWh per 1M generated tokens. Hungriest: B200, 503W, 1.75 kWh. At $0.15/kWh: $0.18 per 1M generated tokens.
Our verdict
Fastest on Olmo-3.1-32B-Think: NVIDIA B300, 87.85 tok/s. Cheapest to rent per job: NVIDIA A100 40GB SXM4, $3.00 per 1M generated tokens.
FAQ
- What GPU do I need to run Olmo-3.1-32B-Think?
- About 20GB. Smallest card that ran it: NVIDIA A10G (24GB).
- How much does it cost to run Olmo-3.1-32B-Think in the cloud?
- $3.00 per 1M generated tokens on a NVIDIA A100 40GB SXM4 at $0.47/hr, cheapest of 9 rentable cards we measured.
- Can I run Olmo-3.1-32B-Think on a 12GB, 16GB or 24GB card?
- It used 18.9GB at the precision we tested. 12GB: no; 16GB: only at Q2_K (16GB); 24GB: yes.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for Olmo-3.1-32B-Think?
- The H100 80GB HBM3: 75.53 vs 46.49 tok/s, 62% faster on our bench.
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.