NVIDIA T4, AI & Machine Learning Benchmarks & Specs

16GB · AI Score 2.9/100 · first-party measured on 12 AI workloads

2.9 AI Score ✓ Measured

Every number on this page is first-party: NVIDIA T4 was run on our pinned 12-workload AI suite on 2026-07-10, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) NVIDIA T4 delivers about 35.01 tokens/sec. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 16GB. For image generation, SDXL runs at 1.18 it/s, while FLUX.1-dev won't fit at BF16 (needs ~26GB). 6 of the 12 workloads won't fit on 16GB at the tested precision, Qwen3 32B, Llama 3.3 70B, FLUX.1-dev, FLUX.1 Kontext and others. We publish those as hard gates rather than quietly dropping to a smaller quant.

AI & Machine Learning benchmark results

Text Generation tok/s 140

SmolLM2-135M416.24

gemma-3-270m362.8

Qwen1.5-0.5B328.29

Qwen2.5-0.5B295.47

Qwen2.5-Coder-0.5B294.96

Qwen3 0.6B263.47

Qwen3-0.6B259.91

LFM2.5-1.2B221.66

Llama 3.2 1B207.71

gemma-3-1b156.88

DeepSeek-R1 Distill 1.5B148.72

Qwen2.5-Coder-1.5B148.51

Workload	Result	Telemetry	Data
SmolLM2-135M	416.24 tok/s	36 W50°CQ4_K_M	✓ Measured
gemma-3-270m	362.8 tok/s	37 W45°CQ4_K_M	✓ Measured
Qwen1.5-0.5B	328.29 tok/s	42 W49°CQ4_K_M	✓ Measured
Qwen2.5-0.5B	295.47 tok/s	43 W44°CQ4_K_M	✓ Measured
Qwen2.5-Coder-0.5B	294.96 tok/s	44 W49°CQ4_K_M	✓ Measured
Qwen3 0.6B	263.47 tok/s	47 W31°CQ4_K_M	✓ Measured
Qwen3-0.6B	259.91 tok/s	49 W45°CQ4_K_M	✓ Measured
Llama 3.2 1B	207.71 tok/s	48 W32°CQ4_K_M	✓ Measured
gemma-3-1b	156.88 tok/s	51 W45°CQ4_K_M	✓ Measured
LFM2.5-1.2B	221.66 tok/s	47 W46°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 1.5B	148.72 tok/s	49 W47°CQ4_K_M	✓ Measured
Qwen2-1.5B	147.68 tok/s	55 W51°CQ4_K_M	✓ Measured
Qwen2.5-1.5B	148.14 tok/s	52 W44°CQ4_K_M	✓ Measured
Qwen2.5-Coder-1.5B	148.51 tok/s	52 W45°CQ4_K_M	✓ Measured
Qwen3 1.7B	139.92 tok/s	57 W34°CQ4_K_M	✓ Measured
Qwen3-1.7B	135.21 tok/s	47 W48°CQ4_K_M	✓ Measured
MiniCPM5 2B	109.06 tok/s	59 W33°CQ4_K_M	✓ Measured
gemma-2-2b	92.13 tok/s	53 W46°CQ4_K_M	✓ Measured
gemma-2-2b-it-abliterated	91.91 tok/s	53 W46°CQ4_K_M	✓ Measured
LFM2.5 2.6B	104.16 tok/s	57 W37°CQ4_K_M	✓ Measured
AI21-Jamba-Reasoning-3B	86.91 tok/s	62 W51°CQ4_K_M	✓ Measured
Granite 4.1 3B	79.3 tok/s	57 W37°CQ4_K_M	✓ Measured
Hermes-3-Llama-3.2-3B	84.99 tok/s	61 W50°CQ4_K_M	✓ Measured
Llama 3.2 3B	85.94 tok/s	54 W47°CQ4_K_M	✓ Measured
Llama-3.2-3B-Instruct-uncensored	86.29 tok/s	56 W48°CQ4_K_M	✓ Measured
Nanbeige4.2-3B	✕ Won't fit needs ~4 GB	VRAM-gated at this precision	✓ Measured
Qwen2.5-3B	86.04 tok/s	55 W45°CQ4_K_M	✓ Measured
Qwen2.5-Coder-3B	85.74 tok/s	59 W45°CQ4_K_M	✓ Measured
SmolLM3 3B	86.08 tok/s	58 W49°CQ4_K_M	✓ Measured
SmolLM3-3B	86.86 tok/s	55 W48°CQ4_K_M	✓ Measured
Phi-4 Mini 3.8B	64.15 tok/s	59 W48°CQ4_K_M	✓ Measured
Gemma 3 4B	65.12 tok/s	60 W47°CQ4_K_M	✓ Measured
Nemotron 3 Nano 4B	65.33 tok/s	62 W36°CQ4_K_M	✓ Measured
Qwen3-4B	67.71 tok/s	53 W45°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	67.45 tok/s	60 W51°CQ4_K_M	✓ Measured
Qwen3-4B-Instruct-2507	66.89 tok/s	61 W52°CQ4_K_M	✓ Measured
Qwen3-4B-Thinking-2507	67 tok/s	62 W52°CQ4_K_M	✓ Measured
phi-2	89.6 tok/s	57 W51°CQ4_K_M	✓ Measured
Dolphin X1 Trinity Nano 6B	114.21 tok/s	54 W49°CQ4_K_M	✓ Measured
Phi-3.5-mini	72.07 tok/s	52 W44°CQ4_K_M	✓ Measured
Phi-4-mini	66 tok/s	59 W46°CQ4_K_M	✓ Measured
DeepSeek Coder 7B Instruct v1.5	46.77 tok/s	62 W39°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 7B	37.58 tok/s	63 W48°CQ4_K_M	✓ Measured
Llama-2-7B	42.98 tok/s	62 W45°CQ4_K_M	✓ Measured
Mistral 7B v0.3	40.01 tok/s	64 W49°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.1	40.63 tok/s	59 W48°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.2	41.31 tok/s	63 W45°CQ4_K_M	✓ Measured
Mistral-7B-Instruct-v0.3	39.87 tok/s	62 W49°CQ4_K_M	✓ Measured
Qwen2.5-7B	39.01 tok/s	64 W46°CQ4_K_M	✓ Measured
Qwen2.5-Coder 7B	37.54 tok/s	63 W49°CQ4_K_M	✓ Measured
Qwen2.5-Coder-7B-Instruct-abliterated	37.31 tok/s	63 W51°CQ4_K_M	✓ Measured
DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored	35.73 tok/s	62 W51°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill Llama 8B	36.36 tok/s	63 W49°CQ4_K_M	✓ Measured
DeepSeek-R1-0528-Qwen3-8B	37.7 tok/s	58 W45°CQ4_K_M	✓ Measured
Dolphin 3.0 Llama 3.1 8B	35.68 tok/s	61 W50°CQ4_K_M	✓ Measured
Dolphin X1 8B	35.66 tok/s	61 W50°CQ4_K_M	✓ Measured
Josiefied-Qwen3-8B-abliterated-v1	35.62 tok/s	62 W51°CQ4_K_M	✓ Measured
L3-8B-Stheno-v3.2	36.28 tok/s	55 W50°CQ4_K_M	✓ Measured
LFM2.5-8B-A1B	139.03 tok/s	55 W51°CQ4_K_M	✓ Measured
Llama 3 8B	41.34 tok/s	63 W41°CQ4_K_M	✓ Measured
Llama 3.1 8B	35.01 tok/s	4.7 GB peak60 W50°C0.59 tok/WQ4_K_M	✓ Measured
Meta-Llama-3.1-8B	37.8 tok/s	61 W45°CQ4_K_M	✓ Measured
Qwen3 8B	36.51 tok/s	61 W48°CQ4_K_M	✓ Measured
Qwen3-8B	37.37 tok/s	62 W46°CQ4_K_M	✓ Measured
dolphin-2.9-llama3-8b	35.1 tok/s	62 W52°CQ4_K_M	✓ Measured
Nemotron Nano 9B v2	31.73 tok/s	62 W40°CQ4_K_M	✓ Measured
Ornith 1.5 9B	34.7 tok/s	62 W40°CQ4_K_M	✓ Measured
Ornith-1.0-9B	32.93 tok/s	61 W46°CQ4_K_M	✓ Measured
gemma-2-9b	28.62 tok/s	62 W48°CQ4_K_M	✓ Measured
Gemma 3 12B	23.12 tok/s	63 W49°CQ4_K_M	✓ Measured
Gemma 3 12B (Q3_K_M)	19.79 tok/s	64 W52°CQ3_K_M	✓ Measured
Gemma 4 12B	24.03 tok/s	62 W50°CQ4_K_M	✓ Measured
NemoMix-Unleashed-12B	23.19 tok/s	63 W53°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 14B	19.21 tok/s	63 W50°CQ4_K_M	✓ Measured
DeepSeek-R1 Distill 14B (Q3_K_M)	16.86 tok/s	64 W53°CQ3_K_M	✓ Measured
EVA-Qwen2.5-14B-v0.2	19.92 tok/s	64 W47°CQ4_K_M	✓ Measured
Hermes-4-14B	19.02 tok/s	64 W53°CQ4_K_M	✓ Measured
Phi-4 14B	17.46 tok/s	63 W50°CQ4_K_M	✓ Measured
Phi-4 14B (Q3_K_M)	16.04 tok/s	65 W52°CQ3_K_M	✓ Measured
Qwen2.5-14B	19.84 tok/s	63 W48°CQ4_K_M	✓ Measured
Qwen2.5-Coder-14B	19.76 tok/s	61 W48°CQ4_K_M	✓ Measured
Qwen2.5-Coder-14B-Instruct-abliterated	18.86 tok/s	65 W52°CQ4_K_M	✓ Measured
Qwen3 14B	19.31 tok/s	64 W50°CQ4_K_M	✓ Measured
Qwen3-14B	19.44 tok/s	63 W50°CQ4_K_M	✓ Measured
Uncensored	18.94 tok/s	65 W53°CQ4_K_M	✓ Measured
StarCoder2 15B	15.66 tok/s	65 W49°CQ4_K_M	✓ Measured
Mistral-Nemo-Instruct-2407	22.98 tok/s	64 W52°CQ4_K_M	✓ Measured
gpt-oss-20b	63.63 tok/s	52 W45°CQ4_K_M	✓ Measured
Codestral 22B	12.59 tok/s	65 W50°CQ4_K_M	✓ Measured
Codestral 22B (Q3_K_M)	11.3 tok/s	66 W50°CQ3_K_M	✓ Measured
GLM-4.7-Flash-REAP-23B-A3B	53.83 tok/s	56 W51°CQ4_K_M	✓ Measured
DeepSeek-Coder-V2-Lite	83.49 tok/s	55 W44°CQ4_K_M	✓ Measured
Cydonia-24B-v4.3	11.21 tok/s	63 W53°CQ4_K_M	✓ Measured
Devstral Small 24B	11.4 tok/s	64 W50°CQ4_K_M	✓ Measured
Dolphin 3.0 R1 Mistral 24B	11.32 tok/s	63 W51°CQ4_K_M	✓ Measured
Dolphin Mistral 24B Venice	11.36 tok/s	65 W51°CQ4_K_M	✓ Measured
Dolphin-Mistral-24B-Venice-Edition	11.33 tok/s	62 W52°CQ4_K_M	✓ Measured
Mistral Small 24B	11.33 tok/s	64 W51°CQ4_K_M	✓ Measured
Mistral Small 24B (Q3_K_M)	10.13 tok/s	66 W54°CQ3_K_M	✓ Measured
Gemma 4 26B A4B	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured
Gemma 3 27B	✕ Won't fit needs ~18 GB	VRAM-gated at this precision	✓ Measured
Qwen3.6 27B	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured
Qwen3.8 27B	✕ Won't fit needs ~21 GB	VRAM-gated at this precision	✓ Measured
Nemotron 3.5 Lightning 30B A3B	✕ Won't fit needs ~30 GB	VRAM-gated at this precision	✓ Measured
Nemotron-3-Nano-30B-A3B	✕ Won't fit needs ~31 GB	VRAM-gated at this precision	✓ Measured
Qwen3 30B A3B	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured
Qwen3 30B A3B (Q3_K_M)	61.82 tok/s	56 W49°CQ3_K_M	✓ Measured
Qwen3 30B A3B Instruct 2507	✕ Won't fit needs ~24 GB	VRAM-gated at this precision	✓ Measured
Qwen3-30B-A3B	✕ Won't fit needs ~24 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Coder 30B A3B	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured
Gemma 4 31B	✕ Won't fit needs ~22 GB	VRAM-gated at this precision	✓ Measured
DeepSeek-R1 Distill 32B	✕ Won't fit needs ~21 GB	VRAM-gated at this precision	✓ Measured
DeepSeek-R1-Distill-Qwen-32B-abliterated	✕ Won't fit needs ~25 GB	VRAM-gated at this precision	✓ Measured
Olmo-3.1-32B-Think	✕ Won't fit needs ~25 GB	VRAM-gated at this precision	✓ Measured
QwQ 32B	✕ Won't fit needs ~21 GB	VRAM-gated at this precision	✓ Measured
Qwen2.5-32B	✕ Won't fit needs ~25 GB	VRAM-gated at this precision	✓ Measured
Qwen2.5-Coder 32B	✕ Won't fit needs ~21 GB	VRAM-gated at this precision	✓ Measured
Qwen2.5-Coder 32B (Q3_K_M)	✕ Won't fit needs ~16 GB	VRAM-gated at this precision	✓ Measured
Qwen3 32B	✕ Won't fit needs ~23 GB	VRAM-gated at this precision	✓ Measured
Dolphin 2.9.1 Yi 1.5 34B	✕ Won't fit needs ~21 GB	VRAM-gated at this precision	✓ Measured
Ornith 1.5 35B A3B	✕ Won't fit needs ~27 GB	VRAM-gated at this precision	✓ Measured
Ornith-1.0-35B	✕ Won't fit needs ~28 GB	VRAM-gated at this precision	✓ Measured
Qwen-AgentWorld-35B-A3B	✕ Won't fit needs ~28 GB	VRAM-gated at this precision	✓ Measured
Qwen3.6 35B A3B	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
GLM-4.7-Flash	✕ Won't fit needs ~23 GB	VRAM-gated at this precision	✓ Measured
Laguna-XS-2.1	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
KAT-Coder-V2.5-Dev	✕ Won't fit needs ~27 GB	VRAM-gated at this precision	✓ Measured
DeepSeek-R1-Distill-Llama-70B	✕ Won't fit needs ~54 GB	VRAM-gated at this precision	✓ Measured
Hermes-4-70B	✕ Won't fit needs ~54 GB	VRAM-gated at this precision	✓ Measured
Llama 3.3 70B	✕ Won't fit needs ~46 GB	VRAM-gated at this precision	✓ Measured
Llama-3.3-70B-Instruct-abliterated	✕ Won't fit needs ~54 GB	VRAM-gated at this precision	✓ Measured
Meta-Llama-3.1-70B	✕ Won't fit needs ~54 GB	VRAM-gated at this precision	✓ Measured
Qwen2.5-72B	✕ Won't fit needs ~60 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Next-80B-A3B	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Next-80B-A3B-Thinking	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Next-80B-A3B-Thinking	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Coder-Next	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Coder-Next	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
Qwen3-Coder-Next-abliterated	✕ Won't fit needs ~61 GB	VRAM-gated at this precision	✓ Measured
gpt-oss-120b	✕ Won't fit needs ~70 GB	VRAM-gated at this precision	✓ Measured

Image Generation images/min 8

Workload	Result	Telemetry	Data
Sana 1.6B	1.13 images/min	69 W46°C	✓ Measured
Stable Diffusion XL	2.36 images/min	12.7 GB peak69 W58°C25.5 s/img	✓ Measured
PixArt-Sigma XL	✕ Won't fit needs ~14 GB	VRAM-gated at this precision	✓ Measured
Stable Diffusion 3.5 Medium	✕ Won't fit needs ~19 GB	VRAM-gated at this precision	✓ Measured
AuraFlow v0.3	✕ Won't fit needs ~24 GB	VRAM-gated at this precision	✓ Measured
FLUX.1 dev	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
Stable Diffusion 3.5 Large	✕ Won't fit needs ~30 GB	VRAM-gated at this precision	✓ Measured
FLUX.1 Schnell	✕ Won't fit needs ~35 GB	VRAM-gated at this precision	✓ Measured

Image Editing images/min 2

Workload	Result	Telemetry	Data
FLUX.1 Kontext dev	✕ Won't fit needs ~26 GB	VRAM-gated at this precision	✓ Measured
Qwen-Image-Edit	✕ Won't fit needs ~42 GB	VRAM-gated at this precision	✓ Measured

Background Removal images/min 1

Workload	Result	Telemetry	Data
BiRefNet	183.48 images/min	52 W30°C	✓ Measured

Upscaling images/min 1

Workload	Result	Telemetry	Data
Swin2SR 4x Upscaler	10.53 images/min	65 W39°C	✓ Measured

Video Generation frames/s 1

Workload	Result	Telemetry	Data
Wan 2.2 5B (720p)	✕ Won't fit needs ~18 GB	VRAM-gated at this precision	✓ Measured

Music Generation x realtime 1

Workload	Result	Telemetry	Data
MusicGen Small	0.92 x realtime	58 W47°C	✓ Measured

Speech to Text x realtime 1

Workload	Result	Telemetry	Data
Whisper large-v3	44.36 x realtime	65 W40°C	✓ Measured

Text to Speech x realtime 1

Workload	Result	Telemetry	Data
Kokoro TTS 82M	42.88 x realtime	51 W38°C	✓ Measured

Depth Estimation images/min 2

Workload	Result	Telemetry	Data
Depth Anything V2 Small	612.25 images/min	27 W28°C	✓ Measured
Depth Anything V2 Large	402.34 images/min	49 W29°C	✓ Measured

Segmentation images/min 2

Workload	Result	Telemetry	Data
SAM ViT-Base	180.73 images/min	67 W31°C	✓ Measured
SAM ViT-Huge	34.91 images/min	70 W35°C	✓ Measured

Vision Language images/min 2

Workload	Result	Telemetry	Data
Florence-2 Base	148.61 images/min	35 W26°C	✓ Measured
Florence-2 Large	83.43 images/min	49 W28°C	✓ Measured

Fine-Tuning train tok/s 4

SmolLM2 1.7B LoRA1561.9

TinyLlama 1.1B LoRA1520.1

Qwen2.5 1.5B LoRA1434.4

Workload	Result	Telemetry	Data
TinyLlama 1.1B LoRA	1520.1 train tok/s	68 W45°C	✓ Measured
Qwen2.5 1.5B LoRA	1434.4 train tok/s	67 W48°C	✓ Measured
SmolLM2 1.7B LoRA	1561.9 train tok/s	69 W50°C	✓ Measured
Qwen2.5 7B LoRA	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured

LLM Serving serve tok/s 4

TinyLlama 1.1B served1897.8

Qwen2.5 1.5B served1337.1

SmolLM2 1.7B served1222.5

Workload	Result	Telemetry	Data
TinyLlama 1.1B served	1897.8 serve tok/s	72 W33°C	✓ Measured
Qwen2.5 1.5B served	1337.1 serve tok/s	70 W36°C	✓ Measured
SmolLM2 1.7B served	1222.5 serve tok/s	70 W37°C	✓ Measured
Qwen2.5 7B served	✕ Won't fit needs ~20 GB	VRAM-gated at this precision	✓ Measured

Speculative Decoding x vs solo 2

Workload	Result	Telemetry	Data
Qwen2.5 1.5B + 0.5B draft	0.9 x vs solo		✓ Measured
Qwen2.5 7B + 0.5B draft	✕ Won't fit needs ~24 GB	VRAM-gated at this precision	✓ Measured

How we measured this. Every result comes from our own pinned, reproducible AI suite, 12 workloads: the Qwen3-4B to Llama-70B LLM ladder (llama.cpp, Q4_K_M), SDXL / Z-Image / FLUX-dev generation, FLUX-Kontext / Qwen-Edit editing, and LTX / Wan video, run first-party on rented hardware with under 0.5% run-to-run variance. Peak VRAM, power draw, temperature and tokens-per-watt are captured per workload. “Won’t fit” rows are real data: where a model exceeds the card’s VRAM at the tested precision we record a hard gate rather than silently dropping to a smaller quant. Measured 2026-07-10 · harness 2.0.0.

NVIDIA T4 specifications

Architecture	Turing
CUDA cores	2,560
VRAM	16GB GDDR6
Memory bus	256-bit
Memory bandwidth	320 GB/s
Boost clock	1,590 MHz
TDP	70 W
Process	12nm
Interface	PCIe 3.0 x16
Release date	2018-09-13
Launch MSRP	$2,299

Verdict, capable, but 16GB sets the ceiling

NVIDIA T4 scores 2.9/100, #61 of 102. It ran 4 of 12; 6 exceeded its 16GB. Every figure here is our own measurement.

Relative performance: where the NVIDIA T4 lands

100% = this card, AI & Machine Learning headline metric (AI Score). #21 of 21 datacenter cards in this vertical.

GPU	%	AI Score
NVIDIA A100 40GB SXM4	586%	17
NVIDIA A100 40GB PCIe	576%	16.7
NVIDIA A10G	221%	6.4
NVIDIA L4	172%	5
NVIDIA T4	100%	2.9

One tier up: NVIDIA L4 (+72%) · compare head-to-head

← All AI & Machine Learning GPU rankings

The silicon

Transistors	13,600 million
Die size	545 mm²
Process node	12 nm
Fabricated by	TSMC
Transistor density	25 million per mm²

Denser than 69% of the 746 cards we have silicon data for. Density is the clearest measure of what a process node bought: a card that gained it without growing the die got its speed from the fab rather than the architecture.

Silicon figures from Wikipedia (CC BY-SA 4.0). Benchmarks on this page are our own. Compare every chip.

What this card can build

Whole-job timings, composed from our measured per-model results on this card.

Workflow	Time	Energy	Basis
Depth pass on a batch	13 s	0.1 Wh	measured
Voiceovers from scripts	55 s	0.59 Wh	measured
Transcribe and subtitle videos	6.9 min	7.28 Wh	measured
Caption a training dataset	12.1 min	9.81 Wh	measured
Masking run	14.4 min	16.59 Wh	measured
Product catalogue cutout	20.3 min	21.65 Wh	all 2 stages measured
Full codebase review	50.7 min	51.37 Wh	measured

Can't run: Product shoot, start to finish (needs FLUX.1 Kontext dev), Product photo shoot (needs FLUX.1 Kontext dev), Photo restoration batch (needs FLUX.1 Kontext dev), Restore and enlarge photos (needs FLUX.1 Kontext dev), Character sheet, 12 poses (needs FLUX.1 dev), Short social clips (needs Qwen3 32B), 60-second AI short film (needs Qwen3 32B), 24-frame storyboard (needs Qwen3 32B), 6-panel comic page (needs Qwen3 32B), Podcast episode pass (needs Qwen3 32B), Long-form article batch (needs Llama 3.3 70B).

Rent or buy?

This card is $2,299 to buy. The cheapest listed rate on Vast.ai is $0.149/hour, but that is the floor: we budget $0.179/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 12,858 GPU-hours. Below it you are paying for idle silicon.

How you would use it	GPU-hours a year	Rental cost a year	Time to break even
2 hours a day, hobby	730	$131	17.6 years
8 hours a day, working on it	2,920	$522	4.4 years
24/7, always-on agent	8,760	$1,566	1.5 years

At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.

Rental price

$0.149/hr+0.0% since 2026-10-01low $0.149 · high $0.150

Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.