NVIDIA · Turing · 16 GB VRAM tier
Best AI Models for NVIDIA T4 (16 GB)
Runs 14B models well; 27–31B barely fits and is slow (14 tok/s).
| # | Model | Speed | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 63.8 t/s | 63.8 tok/s | Q4_K_M · 3.4 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 72.6 t/s | 72.6 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Phi 4Microsoft Q4_K_M · 12.5 GB · 21.8 t/s | 21.8 tok/s | Q4_K_M · 12.5 GB | |
| 5 | Gemma 4 12B ITGoogle Q4_K_M · 9.0 GB · 25.3 t/s | 25.3 tok/s | Q4_K_M · 9.0 GB |
| # | Model | Speed | Coding Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB · 72.6 t/s | 72.6 tok/s | Q4_K_M · 13.1 GB | |
| 2 | Q4_K_M · 7.9 GB · 39.2 t/s | 39.2 tok/s | Q4_K_M · 7.9 GB | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | overall | Q4_K_M · 6.6 GB |
| 4 | Gemma 4 12B ITGoogle Q4_K_M · 9.0 GB · 25.3 t/s | 25.3 tok/s | Q4_K_M · 9.0 GB | |
| 5 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 172 t/s | 172 tok/s | Q4_K_M · 5.5 GB |
| # | Model | Speed | Reasoning Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | Q4_K_M · 6.6 GB | |
| 2 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 72.6 t/s | 72.6 tok/s | Q4_K_M · 13.1 GB | |
| 3 | Qwen3 8BAlibaba Q4_K_M · 7.6 GB · 37.7 t/s | 37.7 tok/s | Q4_K_M · 7.6 GB | |
| 4 | Q4_K_M · 7.9 GB · 39.2 t/s | 39.2 tok/s | Q4_K_M · 7.9 GB | |
| 5 | Gemma 4 12B ITGoogle Q4_K_M · 9.0 GB · 25.3 t/s | 25.3 tok/s | Q4_K_M · 9.0 GB |
| # | Model | Speed | Long context Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 63.8 t/s | 63.8 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Mistral Nemo Instruct 2407Mistral AI Q4_K_M · 11.0 GB · 25.8 t/s | 25.8 tok/s | Q4_K_M · 11.0 GB | |
| 4 | Q4_K_M · 7.9 GB · 39.2 t/s | 39.2 tok/s | Q4_K_M · 7.9 GB | |
| 5 | Gemma 4 12B ITGoogle Q4_K_M · 9.0 GB · 25.3 t/s | 25.3 tok/s | Q4_K_M · 9.0 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | GPT OSS 20BOpenAIHow to run → Q4_K_M · 13.1 GB · 72.6 t/s | 72.6 tok/s | Q4_K_M · 13.1 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 63.8 t/s | 63.8 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | Q4_K_M · 6.6 GB | |
| 4 | Gemma 4 12B ITGoogle Q4_K_M · 9.0 GB · 25.3 t/s | 25.3 tok/s | Q4_K_M · 9.0 GB | |
| 5 | Phi 4Microsoft Q4_K_M · 12.5 GB · 21.8 t/s | 21.8 tok/s | Q4_K_M · 12.5 GB |
| # | Model | Speed | Vision Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.5 9BAlibabaHow to run → Q4_K_M · 6.6 GB · 32.7 t/s | 32.7 tok/s | Q4_K_M · 6.6 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 63.8 t/s | 63.8 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Gemma 3 12B ITGoogle Q4_K_M · 10.9 GB · 25.9 t/s | 25.9 tok/s | Q4_K_M · 10.9 GB | |
| 4 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB · 73.2 t/s | 73.2 tok/s | Q4_K_M · 5.0 GB | |
| 5 | Q4_K_M · 6.5 GB · 33.5 t/s | 33.5 tok/s | Q4_K_M · 6.5 GB |
The T4 runs 14B models well. Turing generation, released 2018.
T4 strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 82.4 tok/s in 15.7 GB
- Low power draw: 70 W board power; a 150 W PSU is enough
Weaknesses
- 70B models don't fit: Llama 3.3 70B Instruct needs 51.3 GB at Q4 vs 16 GB here
- Slow on large dense models: Qwen3.8 27B at 13.9 tok/s
- No FP8 or NVFP4: Turing predates both; use GGUF, AWQ or GPTQ builds
Output speed depends on memory bandwidth (320 GB/s). Grey rows need more than 16 GB.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB32.7tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB21.9tok/s
- 13.9tok/s
- 82.4tok/s
- 70BLlama 3.3 70B Instruct · not at 4-bitneeds 51.3 GB
- 120B MoEGPT OSS 120B · not at 4-bitneeds 70.8 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
All 1057 models on the T4
Featured onlyBest quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
Xgen 7B 8k BaseChat | Q4_K_M | 5.7 GB | 45.0 t/s | 8K | A77, Runs well | |
| BF16 | 7.6 GB | 36.2 t/s | — | A73, Runs well | ||
Qwen3.6 35B A3B Claude 4.7 Opus Reasoning DistilledChatReasoning | Q2_K | 15.7 GB | 82.4 t/s | 262K | C44, Usable | |
DeepSeek Coder v2 Lite BaseChatCode | Q4_K_M | 10.2 GB | 102.3 t/s | 164K | S92, Runs great | |
| Q3_K_L | 15.7 GB | 16.3 t/s | 33K | D28, Barely runs | ||
Yi 6B ChatChat | Q4_K_M | 4.2 GB | 51.1 t/s | 4K | A80, Runs well | |
StarcoderChatCode | Q4_K_M | 13.6 GB | 19.9 t/s | — | B53, Decent | |
Qwen2.5 3BChatFunctions | Q4_K_M | 2.8 GB | 93.3 t/s | 33K | S90, Runs great | |
Qwen 14BChat | Q4_K_M | 10.6 GB | 22.2 t/s | 8K | B62, Decent | |
Falcon 7BChat | Q4_K_M | 7.3 GB | 43.7 t/s | — | A77, Runs well | |
| Q3_K_L | 16.0 GB | 16.0 t/s | 131K | D24, Barely runs | ||
SmolLM 135MChat | Q4_K_M | 0.4 GB | 483.7 t/s | 2K | S100, Runs great | |
Baichuan 7BChat | Q4_K_M | 5.0 GB | 45.0 t/s | 4K | A77, Runs well | |
Gemma 2B ITChat | Q4_K_M | 3.5 GB | 126.1 t/s | — | S95, Runs great | |
| Q4_K_M | 4.6 GB | 45.0 t/s | 2K | A77, Runs well |
Rent NVIDIA T4 in the cloud
Hourly rate on demand for one T4.
Vast.aiMarketplace
from$0.14/hr
Cost per million tokens
Output-token cost on a T4 rented at $0.14/hr, one request at a time. Batch several users on the same card and the cost per token drops several-fold.
| Model | Quant | Speed | Per 1M output |
|---|---|---|---|
| GPT OSS 20B | 73 tok/s | $0.52$0.011 input | |
| Gemma 4 26B A4B IT | 71 tok/s | $0.53$0.011 input | |
| Qwen3.5 9B | 33 tok/s | $1.15$0.025 input | |
| Gemma 4 12B IT | 25 tok/s | $1.49$0.031 input | |
| Qwen3.8 27B | 14 tok/s | $2.71$0.071 input |
Specs
- Architecture
- Turing
- VRAM
- 16 GB GDDR6
- Memory bandwidth
- 320.0 GB/s
- FP16 (dense)
- 65.00 TFLOPS
- CUDA cores
- 2,560
- Tensor cores
- 320
- Compute capability
- 7.5
- TDP
- 70 W
- Released
- 2018-09-12
T4 FAQ
What is the largest model the T4 can run?
Gemma 4 26B A4B IT (25.8B) is the largest at 4-bit or better: at Q4_K_S it uses 15.5 GB of 16 GB, and runs at about 71.1 tok/s. At 3-bit, a bigger model fits too: Gemma 2 27B IT (27.2B) at IQ3_M, 15.0 GB.
Can the T4 run Qwen3.5 9B?
Yes. Qwen3.5 9B fits at Q4_K_M, using 6.6 GB of 16 GB, and it is the best model that runs well on this card.
Can the T4 run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 16 GB on this card.
How many tokens per second does the T4 give?
About 32.7 tok/s on Qwen3.5 9B at Q4_K_M, with the first token after 2.7 s on a 4K-token prompt.
What power supply does the T4 need?
Plan a 150 W PSU (2× the 70 W TDP).
Is 16 GB enough for coding and agents?
Yes for coding: GPT OSS 20B, the best of the 2 coding models that fit, runs at about 73 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.