RTX 3090 24GB vs RTX 4070 Ti Super 16GB for Local LLMs
The RTX 3090 24GB holds bigger models (24.0 GB vs 16.0 GB), and the RTX 3090 24GB generates tokens about 38% faster. Below: speeds on seven popular models, the models only one card can run, and our verdict.
Specs that matter for AI
| RTX 3090 24GB | RTX 4070 Ti Super 16GB | |
|---|---|---|
| Memory | 24 GB | 16 GB |
| GPU-usable memory | 24.0 GB | 16.0 GB |
| Bandwidth | 936 GB/s | 672 GB/s |
| Architecture | Ampere | Ada Lovelace |
| Released | 2020-09 | 2024-01 |
| Llama 3.1 8B Q4_K_M | ~109 tok/s | ~79 tok/s |
Speed and fit by model
Q4_K_M (or native format) with 8K context.
| Model | VRAM | RTX 3090 24GB | RTX 4070 Ti Super 16GB |
|---|---|---|---|
| Llama 3.1 8B | 6.3 GB | Fits ~109 tok/s | Fits ~79 tok/s |
| Qwen 3 14B | 10.5 GB | Fits ~62 tok/s | Fits ~44 tok/s |
| gpt-oss-20b | 13.9 GB | Fits ~155 tok/s | Fits ~124 tok/s |
| Qwen 3 30B-A3B | 18.8 GB | Fits ~114 tok/s | 80% on GPU ~48 tok/s |
| Qwen 3 32B | 21.4 GB | Fits ~28 tok/s | 70% on GPU ~5.5 tok/s |
| Llama 3.3 70B | 43.2 GB | 52% on GPU ~1.8 tok/s | 35% on GPU ~1.4 tok/s |
| gpt-oss-120b | 62.0 GB | 36% on GPU ~20 tok/s | Too small |
Verdict
The used 3090 wins for LLMs. Its 24 GB runs every 27–32B model and the popular 30–35B MoE models at Q4_K_M, which a 16 GB card cannot, and it also has more bandwidth (936 vs 672 GB/s), so it is faster on models that fit both. The 4070 Ti Super is newer, quieter, uses less power and supports FP8 for image generation, but for running language models capacity decides.
Which should you buy?
Choose the RTX 3090 24GB if…
- You want to run 27–35B models.
- You value memory over power efficiency.
- You might add a second card later.
Choose the RTX 4070 Ti Super 16GB if…
- You want a new, quieter, lower-power card.
- Your models are 20B or smaller.
- You generate images and want FP8 support.
Best quality and longest context on each card
The highest-quality quantization that fits fully at 8K context, and the longest context that fits at Q4_K_M with an FP16 cache.
| Model | RTX 3090 24GB: best quant | RTX 4070 Ti Super 16GB: best quant | RTX 3090 24GB: max context | RTX 4070 Ti Super 16GB: max context |
|---|---|---|---|---|
| Qwen 3.5 9B | FP16 / BF16 | Q8_0 | 256K | 256K |
| Qwen 3 14B | Q8_0 | Q6_K | 40K | 37K |
| gpt-oss-20b | MXFP4 (official) | MXFP4 (official) | 128K | 55K |
| Qwen 3.6 27B | Q6_K | IQ4_XS | 98K | – |
| Qwen 3 32B | Q4_K_M | Q2_K | 12K | – |
| Llama 3.3 70B | Does not fit | Does not fit | – | – |
Image generation
Best weight precision for image models with model swapping, at each model's default resolution. Image generation is also limited by compute, so the card with more tensor-core throughput finishes each image faster even when both fit.
| Model | RTX 3090 24GB | RTX 4070 Ti Super 16GB |
|---|---|---|
| SDXL 1.0 | FP16 / BF16 | FP16 / BF16 |
| FLUX.1 [dev] | FP16 / BF16 | FP8 |
| Qwen-Image | FP8 | GGUF Q5_K_S |
| Wan 2.2 A14B (T2V / I2V) | FP8 | GGUF Q5_K_S |
Models only one card can run
Only RTX 3090 24GB
Only RTX 4070 Ti Super 16GB
None
Frequently asked questions
Is the RTX 3090 24GB or RTX 4070 Ti Super 16GB better for local LLMs?
The used 3090 wins for LLMs. Its 24 GB runs every 27–32B model and the popular 30–35B MoE models at Q4_K_M, which a 16 GB card cannot, and it also has more bandwidth (936 vs 672 GB/s), so it is faster on models that fit both. The 4070 Ti Super is newer, quieter, uses less power and supports FP8 for image generation, but for running language models capacity decides.
Which is faster for LLM inference?
The RTX 3090 24GB, by about 38% for token generation on models that fit both (about 109 vs 79 tokens/s on Llama 3.1 8B Q4_K_M).
Which can run bigger models?
The RTX 3090 24GB, with 24.0 GB usable versus 16.0 GB.