T4: 16 GB, what it runs and what it costs
Specs checked · Rental prices from 1h ago · Aliteq Team · How we compute this
The short answer
The NVIDIA Tesla T4 has 16 GB of VRAM and 320 GB/s of memory bandwidth (2018). At Q4_K_M quality or better with an 8K context it runs 17 of the open models we track; the largest that fits with room to spare is DeepSeek R1 Distill 14B. The cheapest rental right now is $0.053/hr (Vast.ai, interruptible).
- VRAM
- 16 GB
- Bandwidth
- 320 GB/s
- Board power
- 70 W
- Launched
- 2018
Source: NVIDIA specification · bandwidth not yet confirmed by a second source, so we don't estimate speed from it
Buy it or rent it
| Provider | Type | From, per hour | Rent |
|---|---|---|---|
| Vast.ai | Spot (can be interrupted) | $0.053/hr | Vast.ai |
| Vast.ai | On demand | $0.054/hr | Vast.ai |
| Lium | Spot (can be interrupted) | $0.224/hr | Lium |
| Lium | On demand | $0.224/hr | Lium |
Lowest price per provider in our hourly snapshot (1h ago). Rent buttons are referral links. Every card side by side: cloud GPU prices.
Which AI models run on the T4
Computed by our VRAM engine for an 8K context: the best quality each model fits at in 16 GB. Speeds are upper-bound estimates from memory bandwidth, not benchmarks.
| Model | Best fit | VRAM used | ~tok/s |
|---|---|---|---|
| Gemma 4 26B-A4B (MoE) | Q4_K_M | 16 GB97% · tight | — |
| Mistral Small 24B | Q4_K_M | 15 GB96% · tight | — |
| gpt-oss 20B (MoE) | Q5_K_M | 15 GB94% · tight | — |
| DeepSeek R1 Distill 14B | Q6_K | 14 GB85% | — |
| Qwen2.5-Coder 14B | Q6_K | 14 GB85% | — |
| Qwen3 14B | Q6_K | 13 GB83% | — |
| Phi-4 14B | Q6_K | 14 GB85% | — |
| Gemma 3 12B | Q8_0 | 13 GB83% | — |
| Gemma 4 12B | Q8_0 | 13 GB80% | — |
| Qwen3.5 9B | Q8_0 | 11 GB66% | — |
| Qwen3 8B | Q8_0 | 10 GB63% | — |
| Gemma 4 E4B | F16 | 16 GB99% · tight | — |
Fits only below Q4 quality: Qwen3.6 35B-A3B (MoE) (Q2_K), Ornith-1.5 35B-A3B (MoE) (Q2_K), Qwen3.5 35B-A3B (MoE) (Q2_K), DeepSeek R1 Distill 32B (Q2_K), Qwen2.5 Coder 32B (Q2_K).
Too big for 16 GB (rent a bigger card instead): what it costs to run Mixtral 8x7B (MoE), what it costs to run DeepSeek R1 Distill 70B, what it costs to run Qwen3-Coder-Next 80B-A3B (MoE), what it costs to run Qwen3-Next 80B-A3B (MoE).
Check a model and context of your own in the VRAM calculator.
AI video on the T4
Alibaba's Wan 2.2 is an open video model you can run yourself in ComfyUI. With 16 GB the T4 fits:
- Wan 2.2 5B, with ComfyUI's native offloading (8 GB and up, per ComfyUI's Wan 2.2 docs)
What each setup needs, the cheapest card to rent and what one clip costs: cloud GPUs for ComfyUI with Flux and Wan 2.2. Against an app or an API: AI video cost per clip, app vs API vs rented GPU. Rather not run a model yourself? Hosted video models such as Seedance, Veo and Hailuo are compared in Kling AI alternatives, priced per clip.
Cards to compare it with
Every card by VRAM tier: all GPUs for gaming and local AI.