NVIDIA RTX 3090 Ti
The NVIDIA RTX 3090 Ti has 24 GB VRAM and 1008 GB/s memory bandwidth. It can run 44 of our 80 tracked models natively in VRAM at 8k context.
With 24 GB GDDR6X, the NVIDIA RTX 3090 Ti is a consumer-tier GPU that can run 44 models natively. It handles 70B-class models at Q4 quantization.
The NVIDIA RTX 3090 Ti is the fully unlocked Ampere consumer flagship with 24GB GDDR6X at 1,008 GB/s — matching the RTX 4090's bandwidth on the older architecture. Its 10,752 CUDA cores deliver slightly faster inference than the base 3090. Shares the same 24GB VRAM tier, handling the same model classes with marginally better tokens-per-second throughput.
NVIDIA RTX 3090 Ti: March 2022 refresh with 24GB at 1008 GB/s — slightly faster than 3090.
Same model compatibility as 3090 with ~10% higher tokens/sec.
Full CUDA support. Rare and expensive — regular 3090 is better value.
| Vendor | NVIDIA |
| Architecture | Ampere |
| VRAM | 24 GB |
| Memory type | GDDR6X |
| Memory bandwidth | 1008 GB/s |
| Compute backend | CUDA |
| Tier | Consumer |
| Released | 2022 |
| Models (native) | 44 / 80 |
| Models (offload) | 7 / 80 |
Software: Full llama.cpp and Ollama support out of the box. CUDA 12.x recommended; driver ≥ 525 required.
Popular models for this GPU
Models this GPU runs natively in VRAM (44)
- Mixtral 8x7B Instruct v0.146.7B · MMLU-Pro 29.7Q2_K · ~37.5 t/s
- Qwen 3.5 35B-A3B (MoE)35B · MMLU-Pro 84.2NVFP4 · ~112.9 t/s
- Qwen 3.6 35B35B · MMLU-Pro 85.2NVFP4 · ~33.3 t/s
- Yi 1.5 34B Chat34.4B · MMLU-Pro 37.0NVFP4 · ~34.1 t/s
- Qwen3 32B32.8B · MMLU-Pro 65.5NVFP4 · ~36.9 t/s
- Qwen 2.5 32B Instruct32.5B · MMLU-Pro 69.0NVFP4 · ~35.6 t/s
- Qwen 2.5 Coder 32B Instruct32.5B · MMLU-Pro 50.4NVFP4 · ~35.6 t/s
- DeepSeek R1 Distill Qwen 32B32.5B · MMLU-Pro 65.0NVFP4 · ~35.6 t/s
- Nemotron 3 Nano 30B32B · MMLU-Pro 78.3NVFP4 · ~120.5 t/s
- Gemma 4 31B31B · MMLU-Pro 85.2NVFP4 · ~35 t/s
- Qwen3 30B-A3B (MoE)30B · MMLU-Pro 61.5NVFP4 · ~112.9 t/s
- Gemma 2 27B Instruct27.2B · MMLU-Pro 38.0NVFP4 · ~39.3 t/s
- Gemma 3 27B Instruct27B · MMLU-Pro 67.5NVFP4 · ~43.6 t/s
- Qwen 3.6 27B27B · MMLU-Pro 86.2NVFP4 · ~43.4 t/s
- Bonsai 27B27B · MMLU-Pro 81.5Ternary (Q2_0) · ~74.4 t/s
- Gemma 4 26B (MoE)26B · MMLU-Pro 82.6NVFP4 · ~84.6 t/s
- Mistral Small 3.1 24B Instruct24B · MMLU-Pro 66.8NVFP4 · ~49.1 t/s
- Mistral Small 22B22.2B · MMLU-Pro 49.2NVFP4 · ~50.5 t/s
- GPT-OSS 20B21B · MMLU-Pro 67.9NVFP4 · ~92.7 t/s
- Qwen3 14B14.8B · MMLU-Pro 61.0NVFP4 · ~74.9 t/s
- Qwen 2.5 14B Instruct14.7B · MMLU-Pro 63.7NVFP4 · ~73.1 t/s
- Phi-4 14B Instruct14B · MMLU-Pro 70.4NVFP4 · ~78.5 t/s
- Mistral Nemo 12B Instruct12.2B · MMLU-Pro 35.6NVFP4 · ~88 t/s
- Gemma 3 12B Instruct12.2B · MMLU-Pro 60.6NVFP4 · ~91.5 t/s
- Gemma 4 12B (Unified)12B · MMLU-Pro 77.2NVFP4 · ~71.1 t/s
- Gemma 2 9B Instruct9.2B · MMLU-Pro 32.0NVFP4 · ~88.3 t/s
- Llama 3.1 8B Instruct8B · MMLU-Pro 48.3BF16 · ~38.4 t/s
- DeepSeek R1 Distill Llama 8B8B · MMLU-Pro 41.0BF16 · ~38.4 t/s
- Qwen3 8B8B · MMLU-Pro 56.7BF16 · ~38.1 t/s
- Qwen 2.5 7B Instruct7.6B · MMLU-Pro 56.3BF16 · ~41.8 t/s
- Mistral 7B Instruct v0.37.25B · MMLU-Pro 30.0BF16 · ~42.1 t/s
- Gemma 3 4B Instruct4B · MMLU-Pro 43.6FP32 · ~39.7 t/s
- Gemma 4 E4B4B · MMLU-Pro 69.4FP32 · ~38.5 t/s
- Phi-3.5 Mini Instruct3.8B · MMLU-Pro 47.4FP32 · ~35.6 t/s
- Llama 3.2 3B Instruct3.2B · MMLU-Pro 24.0FP32 · ~47.7 t/s
- Qwen 2.5 3B Instruct3.1B · MMLU-Pro 32.4FP32 · ~51.6 t/s
- Gemma 2 2B Instruct2.6B · MMLU-Pro 17.8FP32 · ~58.1 t/s
- Gemma 4 E2B2B · MMLU-Pro 60.0FP32 · ~78 t/s
- SmolLM2 1.7B Instruct1.7B · MMLU-Pro 19.0FP32 · ~77.9 t/s
- Qwen 2.5 1.5B Instruct1.5B · MMLU-Pro 16.8FP32 · ~105.1 t/s
- Llama 3.2 1B Instruct1.24B · MMLU-Pro 12.5FP32 · ~125.3 t/s
- Gemma 3 1B Instruct1B · MMLU-Pro 14.7FP32 · ~151.4 t/s
- Qwen 2.5 0.5B Instruct0.5B · MMLU-Pro 10.0FP32 · ~311.9 t/s
- SmolLM2 360M Instruct0.36B · MMLU-Pro 8.0FP32 · ~369 t/s
Models that fit with CPU offload (7)
These use system RAM for layers that don't fit in VRAM — expect much slower inference.
- GLM-4.5 Air 106B106B · MMLU-Pro 81.4Q2_K · ~3.1 t/s
- GLM-4.6V 106B106B · MMLU-Pro 79.9Q2_K · ~3.1 t/s
- Qwen 2.5 72B Instruct72B · MMLU-Pro 71.1NVFP4 · ~1.5 t/s
- Llama 3.3 70B Instruct70B · MMLU-Pro 68.9NVFP4 · ~1.6 t/s
- DeepSeek R1 Distill Llama 70B70B · MMLU-Pro 70.0NVFP4 · ~1.6 t/s
- Llama 3.1 70B Instruct70B · MMLU-Pro 66.4NVFP4 · ~1.6 t/s
- Command-R 35B35B · MMLU-Pro 33.0NVFP4 · ~4 t/s
Too large for this GPU (29)
- Mixtral 8x22B Instruct v0.1
- Llama 3.1 405B Instruct
- DeepSeek V3 671B
- DeepSeek R1 671B
- Llama 4 Scout 109B
- Llama 4 Maverick 400B
- Qwen3 235B-A22B (MoE)
- MiniMax M1 456B
- GPT-OSS 120B
- GLM-4.5 355B
- GLM-4.6 355B
- GLM-4.7 358B
- Qwen 3.5 122B-A10B (MoE)
- MiniMax M2.5 229B
- GLM-5 744B
- MiniMax M2.7 229B
- Nemotron 3 Super 120B
- Kimi K2.6
- GLM-5.1 754B
- DeepSeek V4 Pro 1.6T
- DeepSeek V4 Flash 284B
- GLM-5.2 753B
- Nemotron 3 Ultra 550B-A55B
- Step 3.5 Flash
- Step 3.7 Flash
- MiMo V2.5 Pro
- Kimi K2.5
- MiniMax M3
- Inkling
Compare NVIDIA RTX 3090 Ti with other GPUs
Continue reading
Frequently asked questions
- How much VRAM does the NVIDIA RTX 3090 Ti have?
- The NVIDIA RTX 3090 Ti has 24 GB of GDDR6X with 1008 GB/s memory bandwidth.
- What is the NVIDIA RTX 3090 Ti best for?
- With 24 GB of VRAM, the NVIDIA RTX 3090 Ti is well-suited for running 7B–32B models at Q4 with room for context, making it a great all-rounder for local LLM inference.
- What LLMs can the NVIDIA RTX 3090 Ti run locally?
- The NVIDIA RTX 3090 Ti can run 44 of the 80 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Llama 3.1 8B Instruct at BF16, Llama 3.2 3B Instruct at FP32, Llama 3.2 1B Instruct at FP32.
- Can the NVIDIA RTX 3090 Ti run Llama 3.3 70B Instruct?
- The NVIDIA RTX 3090 Ti can run Llama 3.3 70B Instruct with CPU offload at NVFP4 quantization, but inference will be slower than native VRAM execution.
- Can the NVIDIA RTX 3090 Ti run Qwen 3.6 27B?
- Yes. The NVIDIA RTX 3090 Ti runs Qwen 3.6 27B natively in VRAM at NVFP4 quantization, achieving approximately 43.4 tokens per second.
- Can the NVIDIA RTX 3090 Ti run Llama 3.1 8B Instruct?
- Yes. The NVIDIA RTX 3090 Ti runs Llama 3.1 8B Instruct natively in VRAM at BF16 quantization, achieving approximately 38.4 tokens per second.