Llama-3.2-3B-Instruct-uncensored · 11 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for Llama-3.2-3B-Instruct-uncensored?
Llama-3.2-3B-Instruct-uncensored on 11 GPUs, measured first-party: NVIDIA B300 leads at 448.9 tok/s, T4 trails at 86.3 tok/s, and it peaked at 3GB of VRAM.
Benchmarked weights: bartowski/Llama-3.2-3B-Instruct-uncensored-GGUF
NVIDIA B300
448.9 tok/s on Llama-3.2-3B-Instruct-uncensored, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 448.9 tok/s on Llama-3.2-3B-Instruct-uncensored
- 288GB, clears the Llama-3.2-3B-Instruct-uncensored 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
440.1 tok/s on Llama-3.2-3B-Instruct-uncensored, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 440.1 tok/s on Llama-3.2-3B-Instruct-uncensored
- 96GB, clears the Llama-3.2-3B-Instruct-uncensored 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 Llama-3.2-3B-Instruct-uncensored?, 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.
Llama-3.2-3B-Instruct-uncensored. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA B300 | 448.9 | 12489.9 | 1.39 | 323.4 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 440.1 | 22697.4 | 3.03 | 145.4 W |
| NVIDIA H200 | 429.8 | 16817.9 | 2.32 | 185.6 W |
| NVIDIA H100 80GB HBM3 | 424.2 | 16949.6 | 2.2 | 192.9 W |
| NVIDIA B200 | 419.2 | 17006.1 | 1.31 | 320.3 W |
| NVIDIA L40S | 271 | 18600.6 | 1.76 | 154.3 W |
| NVIDIA A100 80GB SXM4 | 269.1 | 8572.4 | 1.95 | 138.3 W |
| NVIDIA A100 40GB SXM4 | 260.8 | 7097.3 | 2.27 | 114.8 W |
| NVIDIA A10G | 167.6 | 6616.9 | 1.62 | 103.5 W |
| NVIDIA L4 | 101.2 | 6461.2 | 1.82 | 55.5 W |
| NVIDIA T4 | 86.29 | 2565.7 | 1.54 | 56.2 W |
What the numbers show. Across 11 GPUs measured on our own bench, B300 is fastest at 449 tok/s. The slowest, T4, manages 86.3, so the spread is 5.2x from top to bottom. RTX PRO 6000 Blackwell Workstation Edition is the most efficient, 440 tok/s at 145W. Per dollar of launch price, A10G gives the most (59.9 tok/s per $1,000). The fastest card with 16GB or less is T4 at 86.3 tok/s.
How it compares. H100 80GB HBM3: Llama-3.2-3B-Instruct-uncensored 424.2 tok/s, Hermes-3-Llama-3.2-3B 423.7, Llama 3.2 3B 422.9 (3B), MiniCPM5 2B 428.7 (3B), SmolLM3-3B 399.1. 1 of 4 beat Llama-3.2-3B-Instruct-uncensored here.
Cost on a rented GPU. 1M generated tokens of Llama-3.2-3B-Instruct-uncensored: $0.44 on a T4 ($0.14/hr, 3.2 hours), $4.29 on a B300 ($6.94/hr, 37 min, 9.8x the cost).
Llama-3.2-3B-Instruct-uncensored: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA T4 | $0.14/hr | 86.29 | $0.44 |
| NVIDIA A100 40GB SXM4 | $0.47/hr | 260.8 | $0.50 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 440.1 | $0.68 |
| NVIDIA L40S | $0.79/hr | 271 | $0.81 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 269.1 | $0.98 |
| NVIDIA L4 | $0.44/hr | 101.2 | $1.21 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 424.2 | $1.40 |
| NVIDIA H200 | $3.59/hr | 429.8 | $2.32 |
| NVIDIA B200 | $5.98/hr | 419.2 | $3.96 |
| NVIDIA B300 | $6.94/hr | 448.9 | $4.29 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Llama-3.2-3B-Instruct-uncensored. 30+ tok/s: 11 (B300, RTX PRO 6000 Blackwell Workstation Edition, H200). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Llama-3.2-3B-Instruct-uncensored writes anything it reads the input: 22697.4 tok/s on the RTX PRO 6000 Blackwell Workstation Edition (0.2s for a 4,000-token prompt), 2565.7 on the T4 (1.6s). Long documents and big code files feel this number more than the generation speed.
VRAM for Llama-3.2-3B-Instruct-uncensored. Measured peak 2.7GB, so 8GB is the smallest common card size; smallest card it ran on: T4 (16GB). With long context: Q4_K_M 3GB (tested), Q2_K 2GB, Q3_K_M 3GB, Q5_K_M 4GB, Q6_K 4GB.
Power on Llama-3.2-3B-Instruct-uncensored. Most efficient: RTX PRO 6000 Blackwell Workstation Edition, 145W, 91.8 Wh per 1M generated tokens. Hungriest: B300, 323W, 0.20 kWh. At $0.15/kWh: $0.014 per 1M generated tokens.
Our verdict
Fastest on Llama-3.2-3B-Instruct-uncensored: NVIDIA B300, 448.9 tok/s. Cheapest to rent per job: NVIDIA T4, $0.44 per 1M generated tokens.
FAQ
- What GPU do I need to run Llama-3.2-3B-Instruct-uncensored?
- About 3GB. Smallest card that ran it: NVIDIA T4 (16GB).
- How much does it cost to run Llama-3.2-3B-Instruct-uncensored in the cloud?
- $0.44 per 1M generated tokens on a NVIDIA T4 at $0.14/hr, cheapest of 10 rentable cards we measured.
- Can I run Llama-3.2-3B-Instruct-uncensored on a 12GB, 16GB or 24GB card?
- It used 2.7GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for Llama-3.2-3B-Instruct-uncensored?
- The H100 80GB HBM3: 424.2 vs 269.1 tok/s, 58% 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.