NVIDIA · Ada Lovelace · 24 GB VRAM tier
Best AI Models for NVIDIA L4 (24 GB)
Runs 14B and 35B MoE models well; 27–31B fits but is slow (11 tok/s).
24 GB GDDR6300 GB/s121 TFLOPS FP1672 WAda Lovelace 2023
| # | Model | Speed | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 22.0 GB | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB | |
| 3 | Qwen3 30B A3B Thinking 2507Alibaba Q4_K_M · 19.4 GB · 75.9 t/s | 75.9 tok/s | Q4_K_M · 19.4 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 68.5 t/s | 68.5 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Q4_K_M · 19.3 GB · 97.3 t/s | 97.3 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Coding Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 22.0 GB | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB | |
| 3 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 68.5 t/s | 68.5 tok/s | Q4_K_M · 13.1 GB | |
| 4 | Q4_K_M · 7.9 GB · 36.8 t/s | 36.8 tok/s | Q4_K_M · 7.9 GB | |
| 5 | Q4_K_M · 19.3 GB · 97.3 t/s | 97.3 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Reasoning Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 22.0 GB | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.7 t/s | 30.7 tok/s | Q4_K_M · 6.6 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 68.5 t/s | 68.5 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Q4_K_M · 19.3 GB · 97.3 t/s | 97.3 tok/s | Q4_K_M · 19.3 GB |
| # | Model | Speed | Math Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 22.0 GB | |
| 2 | Qwen3.8 27BAlibaba Q4_K_M · 17.9 GB · 11.2 t/s | 11.2 tok/s | Q4_K_M · 17.9 GB | |
| 3 | Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 11.2 t/s | 11.2 tok/s | Q4_K_M · 17.9 GB | |
| 4 | Q4_K_M · 19.3 GB · 97.3 t/s | 97.3 tok/s | Q4_K_M · 19.3 GB | |
| 5 | GLM 4.7 FlashZ.ai Q4_K_M · 19.9 GB · 56.9 t/s | 56.9 tok/s | Q4_K_M · 19.9 GB |
| # | Model | Speed | Long context Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB | |
| 2 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.7 t/s | 30.7 tok/s | Q4_K_M · 6.6 GB | |
| 3 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 59.8 t/s | 59.8 tok/s | Q4_K_M · 3.4 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 68.5 t/s | 68.5 tok/s | Q4_K_M · 13.1 GB | |
| 5 | LFM2.5 24B A2BLiquid AI Q4_K_M · 15.0 GB · 100 t/s | 100 tok/s | Q4_K_M · 15.0 GB |
| # | Model | Speed ↓ | llmrun Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Q4_K_M · 19.4 GB · 75.9 t/s | 75.9 tok/s | Q4_K_M · 19.4 GB | |
| 2 | Q4_K_M · 19.3 GB · 97.3 t/s | 97.3 tok/s | Q4_K_M · 19.3 GB | |
| 3 | Qwen3 30B A3B Instruct 2507Alibaba Q4_K_M · 19.4 GB · 75.9 t/s | 75.9 tok/s | Q4_K_M · 19.4 GB | |
| 4 | GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 68.5 t/s | 68.5 tok/s | Q4_K_M · 13.1 GB | |
| 5 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB |
| # | Model | Speed | Vision Score | Quant · VRAM |
|---|---|---|---|---|
| 1 | Qwen3.6 35B A3BAlibabaHow to run → Q4_K_M · 22.0 GB · 58.6 t/s | 58.6 tok/s | Q4_K_M · 22.0 GB | |
| 2 | Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 63.5 t/s | 63.5 tok/s | Q4_K_M · 16.5 GB | |
| 3 | Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 30.7 t/s | 30.7 tok/s | Q4_K_M · 6.6 GB | |
| 4 | Gemma 3 12B ITGoogle Q4_K_M · 10.9 GB · 24.3 t/s | 24.3 tok/s | Q4_K_M · 10.9 GB | |
| 5 | MiniCPM V 4.7 35B A3BOpenBMB Q4_K_M · 22.1 GB · 67.8 t/s | 67.8 tok/s | Q4_K_M · 22.1 GB |
See all 327 models on the L4 →Best quant that fits · speeds estimated on this card
The L4 runs 14B and 35B MoE models well. Ada Lovelace generation, released 2023.
GenerationAda Lovelacereleased 2023
L4 strengths and weaknesses
Strengths
- Runs 35B MoE models: Qwen3.6 35B A3B at 58.6 tok/s in 22 GB
- Reads long prompts fast: A 4K-token prompt on Qwen3.6 35B A3B starts answering in 714 ms
- Low power draw: 72 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 24 GB here
- Slow on large dense models: Qwen3.8 27B at 11.2 tok/s
Output speed depends on memory bandwidth (300 GB/s). Grey rows need more than 24 GB.
- 9BQwen3.5 9B · Q4_K_M · 6.6 GB30.7tok/s
- 14BQwen3 14B · Q4_K_M · 11.9 GB20.5tok/s
- 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB11.2tok/s
- 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB58.6tok/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 1141 models on the L4
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 |
|---|---|---|---|---|---|---|
Pythia 160MChat | Q4_K_M | 0.1 GB | 1392.9 t/s | 2K | S100, Runs great | |
Qwen2.5 VL 3B InstructVision | Q4_K_M | 3.2 GB | 74.1 t/s | 128K | S86, Runs great | |
Sarvam 1Chat | Q4_K_M | 2.8 GB | 95.1 t/s | 8K | S90, Runs great | |
| Q4_K_M | 6.2 GB | 36.2 t/s | 8K | A73, Runs well | ||
Qwen3 1.7BChatFunctions | Q4_K_M | 3.4 GB | 111.4 t/s | 41K | S94, Runs great | |
Qwen3 VL 32B InstructVisionFunctions | Q4_K_M | 23 GB | 9.4 t/s | 262K | D23, Barely runs | |
Tiny Random Phi3Chat | Q4_K_M | 0.3 GB | 650.0 t/s | 4K | S100, Runs great | |
C4ai Command R7b 12 2024ChatFunctions | Q4_K_M | 7.9 GB | 36.8 t/s | — | A74, Runs well | |
Mellum2 12B A2.5B InstructChatFunctions | Q4_K_M | 7.8 GB | 106.1 t/s | 131K | S92, Runs great | |
Gemma 2 9B ITChat | Q4_K_M | 7.3 GB | 32.0 t/s | 8K | A71, Runs well | |
| Q4_K_M | 1.7 GB | 224.1 t/s | 33K | S100, Runs great | ||
| Q4_K_M | 17.8 GB | 13.1 t/s | 33K | B52, Decent | ||
| BF16 | 7.1 GB | 31.0 t/s | 8K | A71, Runs well | ||
| Q4_K_M | 4.1 GB | 79.3 t/s | 131K | S87, Runs great | ||
| Q4_K_M | 3.7 GB | 64.6 t/s | 8K | A84, Runs well |
Rent NVIDIA L4 in the cloud
Hourly rate on demand for one L4.
Cost per million tokens
Output-token cost on an L4 rented at $0.44/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 |
|---|---|---|---|
| NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 | 97 tok/s | $1.26$0.011 input | |
| Gemma 4 26B A4B IT | 64 tok/s | $1.92$0.020 input | |
| Qwen3.6 35B A3B | 59 tok/s | $2.09$0.021 input | |
| Qwen3.8 27B | 11 tok/s | $10.91$0.12 input | |
| Muse Glimmer 30B | 11 tok/s | $11.53$0.13 input |
Specs
- Architecture
- Ada Lovelace
- VRAM
- 24 GB GDDR6
- Memory bandwidth
- 300.0 GB/s
- FP16 (dense)
- 121.00 TFLOPS
- CUDA cores
- 7,424
- Tensor cores
- 240
- Compute capability
- 8.9
- TDP
- 72 W
- Released
- 2023-03-21
L4 FAQ
What is the largest model the L4 can run?
Qwen3.6 35B A3B (36.0B) is the largest at 4-bit or better: at Q4_K_M it uses 22.0 GB of 24 GB, and runs at about 58.6 tok/s.
Can the L4 run 70B models?
No. Llama 3.3 70B Instruct needs 51.3 GB at Q4, more than the 24 GB on this card.
How many tokens per second does the L4 give?
About 58.6 tok/s on Qwen3.6 35B A3B at Q4_K_M, with the first token after 0.7 s on a 4K-token prompt.
What power supply does the L4 need?
Plan a 150 W PSU (2× the 72 W TDP).
Is 24 GB enough for coding and agents?
Yes for coding: Qwen3.6 35B A3B, the best of the 4 coding models that fit, runs at about 59 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.