What AI models run on the RTX 3090 Ti?
The RTX 3090 Ti has 24 GB of VRAM (1,008 GB/s). About 22.8 GB is usable for a model. 209 of 298 open models we track fit at 8K context.
This card has 24 GB of VRAM. See every model that fits in 24 GB VRAM. Other 24 GB hardware: RTX 3090, RTX 4090, RX 7900 XTX, Mac M2 24GB, Mac M3 24GB, Mac M4 Pro 24GB.
Compare for local LLMs: RTX 3090 Ti vs RTX 3090 all comparisons
Using a specific tool? Ollama models that run on the RTX 3090 Ti · LM Studio · llama.cpp offload
Models that fit, largest first
Bigger models with CPU offload
These popular models don't fit in the RTX 3090 Ti's VRAM at 4-bit, but run with llama.cpp offload, keeping part of the model in system RAM. Speed range is DDR4-3200 to DDR5-6000 (estimate). Method.
| Model | Params | Offload plan |
|---|---|---|
| GLM 5.3 Flash | 321B MoE | Offload Q2_K, 128 GB RAM · ~8.1–16 tok/s |
| gpt-oss 120B | 117B MoE | Offload Q4_K_M, 64 GB RAM · ~17–34 tok/s |
| DeepSeek V4 Flash 0731 | 304B MoE | Offload Q3_K_M, 128 GB RAM · ~6.8–13 tok/s |
| DeepSeek V4 Flash | 291B MoE | Offload Q3_K_M, 128 GB RAM · ~7.7–15 tok/s |
| GLM 4.5 Air | 110B MoE | Offload Q4_K_M, 64 GB RAM · ~6.5–13 tok/s |
| MiniMax M2.5 | 229B MoE | Offload Q4_K_M, 128 GB RAM · ~6.1–12 tok/s |
| Hy3 | 299B MoE | Offload Q3_K_M, 128 GB RAM · ~4.5–9.1 tok/s |
| Nex N2.6 | 321B MoE | Offload Q2_K, 128 GB RAM · ~7.5–15 tok/s |
Comparing cards? The RTX 3090 Ti (launch MSRP $1999) is in best GPUs for local LLMs under $2,000. See also Mac vs PC.
Just too big at 4-bit
These need up to 35% more memory than the RTX 3090 Ti has at Q4_K_M and 8K context. A smaller quant or shorter context may work.
| Model | Q4_K_M needs | Short by |
|---|---|---|
| Hermes 4.3 36B | 23.0 GB | 0.2 GB |
| Seed OSS 36B Instruct | 23.0 GB | 0.2 GB |
| Flint | 24.5 GB | 1.7 GB |
| Falcon 40B | 25.2 GB | 2.4 GB |
| Nous Hermes 2 Mixtral 8x7B DPO | 28.0 GB | 5.2 GB |
| Kimi Linear 48B A3B Instruct | 28.9 GB | 6.1 GB |
Questions
What is the biggest model that runs on the RTX 3090 Ti?
Of the models we track, the largest that fits at 8K context is Kimi Linear 48B A3B Instruct (49B) at Q2_K, needing about 17.3 GB.
How much of the RTX 3090 Ti's memory can a model use?
About 22.8 GB: we assume 95% of the 24 GB is usable after driver and display overhead.
Spec source: NVIDIA specifications. Model data: Hugging Face Hub API. Speed is a bandwidth ceiling (bandwidth ÷ bytes read per token), not a benchmark. Methodology · Report an error.