SmolLM2-360M · 4 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for SmolLM2-360M?
SmolLM2-360M on 4 GPUs, measured first-party: NVIDIA H100 80GB HBM3 leads at 787.6 tok/s, L4 trails at 440 tok/s, and it peaked at 2GB of VRAM.
Benchmarked weights: bartowski/SmolLM2-360M-Instruct-GGUF
NVIDIA H100 80GB HBM3
787.6 tok/s on SmolLM2-360M, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 787.6 tok/s on SmolLM2-360M
- 80GB, clears the SmolLM2-360M floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
What GPU Do You Need for SmolLM2-360M?, 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.
SmolLM2-360M. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 787.6 | 33002.3 | 5.69 | 138.4 W |
| NVIDIA L40S | 714.5 | 43427.2 | 7.93 | 90.1 W |
| NVIDIA A10G | 491.8 | 23935.2 | 7.26 | 67.7 W |
| NVIDIA L4 | 440.2 | 25496 | 13.14 | 33.5 W |
What the numbers show. Across 4 GPUs measured on our own bench, H100 80GB HBM3 is fastest at 788 tok/s. The slowest, L4, manages 440, so the spread is 1.8x from top to bottom. L4 is the most efficient, 440 tok/s at 34W. Per dollar of launch price, L4 gives the most (176.1 tok/s per $1,000).
About SmolLM2-360M. SmolLM2-360M: from HuggingFaceTB, 0.4B parameters, on Hugging Face since October 2024, Apache 2.0 licence. 650,037 downloads in the last 30 days and 1 community quantizations.
How it compares. H100 80GB HBM3: SmolLM2-360M 787.6 tok/s, Qwen2.5-0.5B 890.0, Qwen2 0.5B 881.9, Qwen3 0.6B 713.5, gemma-3-1b 504.2 (1B). 2 of 4 beat SmolLM2-360M here.
Cost on a rented GPU. 1M generated tokens of SmolLM2-360M: $0.28 on a L4 ($0.44/hr, 38 min), $0.75 on a H100 80GB HBM3 ($2.14/hr, 21 min, 2.7x the cost).
SmolLM2-360M: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA L4 | $0.44/hr | 440.2 | $0.28 |
| NVIDIA L40S | $0.79/hr | 714.5 | $0.31 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 787.6 | $0.75 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for SmolLM2-360M. 30+ tok/s: 4 (H100 80GB HBM3, L40S, A10G). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before SmolLM2-360M writes anything it reads the input: 43427.2 tok/s on the L40S (0.1s for a 4,000-token prompt), 23935.2 on the A10G (0.2s). Long documents and big code files feel this number more than the generation speed.
VRAM for SmolLM2-360M. Measured peak 1.0GB, so 8GB is the smallest common card size; smallest card it ran on: A10G (24GB). With long context: Q4_K_M 1GB (tested), Q2_K 1GB, Q3_K_M 1GB, Q5_K_M 1GB, Q6_K 1GB.
Power on SmolLM2-360M. Most efficient: L4, 34W, 21.1 Wh per 1M generated tokens. Hungriest: L40S, 90W, 35.0 Wh.
Our verdict
Fastest on SmolLM2-360M: NVIDIA H100 80GB HBM3, 787.6 tok/s. Cheapest to rent per job: NVIDIA L4, $0.28 per 1M generated tokens.
FAQ
- What GPU do I need to run SmolLM2-360M?
- About 1GB. Smallest card that ran it: NVIDIA A10G (24GB).
- How much does it cost to run SmolLM2-360M in the cloud?
- $0.28 per 1M generated tokens on a NVIDIA L4 at $0.44/hr, cheapest of 3 rentable cards we measured.
- Can I run SmolLM2-360M on a 12GB, 16GB or 24GB card?
- It used 1.0GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the H100 80GB HBM3 or the L40S faster for SmolLM2-360M?
- The H100 80GB HBM3: 787.6 vs 714.5 tok/s, 10% 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.