# GeForce RTX 3090

> OnPrem World · https://onpremworld.com/en/hardware/rtx-3090 · Last updated: 2026-10-14 · CC BY 4.0

24 GB of VRAM, the same as the RTX 4090, with 93% of its memory bandwidth, so it runs 27–32B models at 4-bit. NVIDIA also lists NVLink for pairing two cards.

## Specifications

|  |  |
|---|---|
| Memory | 24 GB GDDR6X |
| Memory bandwidth | 936 GB/s |
| Power | 350 W |
| Launch | 2020-09 |
| Launch price | US$1,499 |
| Architecture | Ampere |
| Software | CUDA |

## What it can run (24 GB, 8K)

Estimates, not benchmarks. See the methodology.

| Model | Parameters | Best quant | Total | tok/s (est.) |
|---|---|---|---|---|
| Qwen3.5-35B-A3B | 35B | Q4_K_M | 21.4 GiB | 360 |
| Qwen3.6-35B-A3B | 35B | Q4_K_M | 21.4 GiB | 360 |
| Qwen3-32B | 32.8B | Q4_K_M | 21.9 GiB | 33 |
| DeepSeek-R1-Distill-Qwen-32B | 32.76B | Q4_K_M | 21.9 GiB | 33 |
| Gemma 4 31B IT | 30.7B | Q5_K_M | 22.9 GiB | 30 |
| Qwen3-30B-A3B (2507) | 30.5B | Q5_K_M | 22.5 GiB | 279 |
| Qwen3-Coder-30B-A3B-Instruct | 30.5B | Q5_K_M | 22.5 GiB | 279 |
| GLM-4.7-Flash (30B-A3B) | 30B | Q5_K_M | 21.8 GiB | 307 |
| Qwen3.5-27B | 27B | Q6_K | 22.7 GiB | 29 |
| Qwen3.6-27B | 27B | Q6_K | 22.7 GiB | 29 |
| Qwen3.8-27B | 27B | Q6_K | 22.7 GiB | 29 |
| Gemma 3 27B IT | 27B | Q6_K | 23.3 GiB | 29 |
| Gemma 4 26B A4B IT | 25.2B | Q6_K | 21.1 GiB | 209 |
| Mistral Small 3.2 24B Instruct (2506) | 24B | Q6_K | 21.1 GiB | 33 |
| Magistral Small 1.2 (2509) | 24B | Q6_K | 21.1 GiB | 33 |
| Devstral Small 2 24B Instruct (2512) | 24B | Q6_K | 21.1 GiB | 33 |
| gpt-oss-20b | 21B | Q8_0 | 22.5 GiB | 171 |
| Phi-4 (14B) | 14B | Q8_0 | 16.6 GiB | 44 |
| Gemma 3 12B IT | 12B | Q8_0 | 13.8 GiB | 51 |
| Gemma 4 12B Unified IT | 11.95B | Q8_0 | 13.3 GiB | 52 |
| Qwen3.5-9B | 9B | FP16 | 18.4 GiB | 36 |
| Qwen3-8B | 8.2B | FP16 | 17.7 GiB | 40 |
| DeepSeek-R1-0528-Qwen3-8B | 8.19B | FP16 | 17.6 GiB | 40 |
| Llama 3.1 8B Instruct | 8B | FP16 | 17.1 GiB | 41 |
| Gemma 4 E4B IT | 8B | FP16 | 16.4 GiB | 41 |
| Llama 3.2 3B Instruct | 3.21B | FP16 | 7.7 GiB | 102 |
| Kimi K2.6 | 1000B | Doesn't fit | – | – |
| DeepSeek-R1-0528 (671B) | 671B | Doesn't fit | – | – |
| DeepSeek-V4-Flash | 284B | Doesn't fit | – | – |
| Qwen3.8-Flash-Next | 180B | Doesn't fit | – | – |
| Qwen3.5-122B-A10B | 122B | Doesn't fit | – | – |
| Mistral Small 4 119B (2603) | 119B | Doesn't fit | – | – |
| gpt-oss-120b | 117B | Doesn't fit | – | – |
| Llama 4 Scout 17B-16E Instruct | 109B | Doesn't fit | – | – |
| GLM-4.5-Air | 106B | Doesn't fit | – | – |
| Qwen3-Coder-Next (80B-A3B) | 80B | Doesn't fit | – | – |
| Llama 3.3 70B Instruct | 70B | Doesn't fit | – | – |

## Prices

No prices recorded for this device yet.

## Frequently asked questions

### Which LLMs can the GeForce RTX 3090 run?

With 24 GB and an 8K context, the largest models that fit include: Qwen3.5-35B-A3B, Qwen3.6-35B-A3B, Qwen3-32B, DeepSeek-R1-Distill-Qwen-32B, Gemma 4 31B IT. See the full table below.

### How fast is the GeForce RTX 3090 for local LLMs?

Generation speed is mostly set by memory bandwidth (936 GB/s). For Llama 3.1 8B Instruct at FP16, expect roughly 41 tokens/s; bigger models are slower.

### Can the GeForce RTX 3090 run 70B models?

Llama 3.3 70B Instruct does not fit on the GeForce RTX 3090 (24 GB), even at Q4_K_M with an 8K context.

## Sources

- https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/
- https://www.nvidia.com/content/PDF/nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf
- https://nvidianews.nvidia.com/news/nvidia-delivers-greatest-ever-generational-leap-in-performance-with-geforce-rtx-30-series-gpus
- https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units