NVIDIA - consumer GPU - verified
Local LLMs that run on NVIDIA GeForce RTX 5090 (32 GB)
By this directory's rule of thumb, 30 of 63 open-weight models fit in the 32 GB of VRAM on the RTX 5090. The largest by total parameters is NVIDIA Nemotron 3.5 Bolt 74B-A9B (74.9B), whose smallest non-MLX download is 22.0 GB, the 24GB (native 2-bit) GGUF build from nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF.
- Models that fit
- 30 of 63
- Memory
- 32 GB
- Memory type
- GDDR7
- Bandwidth
- 1,792 GB/s
Spec sheet
- Memory
- 32 GB
- Memory type
- GDDR7
- Unified memory
- No
- Bandwidth
- 1,792 GB/s
- Released
- 2025
512-bit memory interface; NVIDIA states bandwidth as 1792 GB/sec.
Source: NVIDIA spec page (nvidia.com)
30 models that fit in 32 GB
Total params, largest first| Model | Total params | Smallest download | Quant | Commercial |
|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Bolt 74B-A9BNVIDIA | 74.9B9B active | 22.0 GB | 24GB (native 2-bit)GGUF | Commercial use: Yes |
| Qwen3.6-35B-A3BAlibaba Cloud (Qwen) | 36B3B active | 22.1 GB | UD-Q4_K_MGGUF | Commercial use: Yes |
| Cohere Aya Expanse 32BCohere | 32.3B | 19.8 GB | Q4_K_MGGUF | Commercial use: No |
| Ai2 Olmo 3.1 32B ThinkAi2 | 32.2B | 19.5 GB | Q4_K_MGGUF | Commercial use: Yes |
| NVIDIA Nemotron 3.5 Lightning 30B-A3BNVIDIA | 31.6B3B active | 18.9 GB | Q4_0GGUF | Commercial use: Yes |
| Gemma 4 31BGoogle | 31.3B | 17.7 GB | Q4_0 (QAT)GGUF | Commercial use: Yes |
| IBM Granite 4.2 30BIBM | 29.3B | 17.7 GB | Q4_K_MGGUF | Commercial use: Yes |
| Qwen3.8-27BAlibaba Cloud (Qwen) | 27.8B | 16.5 GB | UD-Q4_K_MGGUF | Commercial use: Yes |
| Gemma 4 26B A4BGoogle | 25.8B3.8B active | 14.4 GB | Q4_0 (QAT)GGUF | Commercial use: Yes |
| Devstral Small 2 24B Instruct 2512Mistral AI | 24B | 14.3 GB | Q4_K_MGGUF | Commercial use: Yes |
| Liquid AI LFM2.5-24B-A2BLiquid AI | 24B2.5B active | 14.5 GB | Q4_K_MGGUF | Commercial use: Conditions |
| gpt-oss-20bOpenAI | 20.9B3.6B active | 11.6 GB | Q4_K_MGGUF | Commercial use: Yes |
| Phi-4Microsoft | 14.7B | 9.1 GB | Q4_KGGUF | Commercial use: Yes |
| Phi-4-reasoning-plusMicrosoft | 14.7B | 9.1 GB | Q4_K_MGGUF | Commercial use: Yes |
| Ministral 3 14B Instruct 2512Mistral AI | 13.9B | 8.2 GB | Q4_K_MGGUF | Commercial use: Yes |
| Gemma 4 12BGoogle | 12B | 7.0 GB | Q4_0 (QAT)GGUF | Commercial use: Yes |
| Qwen3.5-9BAlibaba Cloud (Qwen) | 9.7B | 5.7 GB | Q4_K_MGGUF | Commercial use: Yes |
| Ministral 3 8B Instruct 2512Mistral AI | 8.9B | 5.2 GB | Q4_K_MGGUF | Commercial use: Yes |
| IBM Granite 4.2 8BIBM | 8.8B | 5.3 GB | Q4_K_MGGUF | Commercial use: Yes |
| Cohere Command R7B (12-2024)Cohere | 8B | 5.1 GB | Q4_K_MGGUF | Commercial use: No |
| Gemma 4 E4BGoogle | 8B | 5.0 GB | Q4_K_MGGUF | Commercial use: Yes |
| Ai2 Olmo Hybrid Instruct DPO 7BAi2 | 7.4B | 14.9 GB | BF16Official weights | Commercial use: Yes |
| Gemma 4 E2BGoogle | 5.1B | 3.1 GB | Q4_K_MGGUF | Commercial use: Yes |
| Qwen3.5-4BAlibaba Cloud (Qwen) | 4.7B | 2.7 GB | Q4_K_MGGUF | Commercial use: Yes |
| Ministral 3 3B Instruct 2512Mistral AI | 3.8B | 2.1 GB | Q4_K_MGGUF | Commercial use: Yes |
| Phi-4-mini-instructMicrosoft | 3.8B | 2.5 GB | Q4_K_MGGUF | Commercial use: Yes |
| Phi-4-mini-reasoningMicrosoft | 3.8B | 2.5 GB | Q4_K_MGGUF | Commercial use: Yes |
| Liquid AI LFM2.5-2.6BLiquid AI | 2.7B | 1.7 GB | Q4_K_MGGUF | Commercial use: Conditions |
| Qwen3.5-2BAlibaba Cloud (Qwen) | 2.3B | 1.3 GB | Q4_K_MGGUF | Commercial use: Yes |
| Qwen3.5-0.8BAlibaba Cloud (Qwen) | 0.9B | 0.5 GB | Q4_K_MGGUF | Commercial use: Yes |
The rest of the 63 need more than 32 GB. Browse every local model.
Other devices with 32 GB
Same memory size. Each shows how many models fit under its own rule (Macs also count MLX builds), and the memory type and bandwidth differ.
Compare all 33 GPUs and Macs by memory, bandwidth and how many models fit.
FAQ
How many local LLMs can the RTX 5090 run?
30 of the 63 open-weight models in this directory fit, with 32 GB of VRAM, by this site's rule of thumb that the smallest non-MLX download is at most 85% of the memory. MLX builds do not count here: they run on Apple silicon only.
Can the RTX 5090 run NVIDIA Nemotron 3.5 Bolt 74B-A9B?
Yes, by this directory's rule of thumb: its smallest non-MLX download, 22.0 GB, is at most 85% of 32 GB (27.2 GB), the 24GB (native 2-bit) GGUF build from nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF. That leaves room for the context. It is a memory check, not a speed estimate.
Can the RTX 5090 run Liquid AI LFM2.5-24B-A2B?
Yes, by this directory's rule of thumb: its smallest non-MLX download, 14.5 GB, is at most 85% of 32 GB (27.2 GB), the Q4_K_M GGUF build from LiquidAI/LFM2.5-24B-A2B-GGUF. That leaves room for the context. It is a memory check, not a speed estimate.
Can the RTX 5090 run Phi-4-reasoning-vision-15B?
Not by this directory's rule of thumb: its smallest non-MLX download, 30.2 GB, is more than 85% of 32 GB (27.2 GB). It needs the 48 GB tier or more: https://www.developersdigest.tech/local-models/fits/48.
Can the RTX 5090 use system RAM for bigger models?
This page only counts models that fit entirely in the 32 GB of VRAM. Runtimes can offload layers to system RAM, but this directory does not list or measure that.
What memory bandwidth does the RTX 5090 have?
NVIDIA states 1,792 GB/s of memory bandwidth (GDDR7). From the spec notes: 512-bit memory interface; NVIDIA states bandwidth as 1792 GB/sec. This page does not turn bandwidth into a speed estimate.
Newsletter
New open weights, every week
One email per week. New model releases, open-weight drops and what they mean for what you build.
Keep exploring
Best local AI models
Our curated picks by hardware tier, with benchmarks and the reasoning.
Model API pricing and specs
First-party prices per 1M tokens when you would rather call an API.
AI benchmark leaderboards
Sourced scores, for when you need to compare quality.
Setup guides
Point your coding tools at the model you picked.