What can NVIDIA GeForce RTX 3090 run for reasoning?

Build: RTX 3090 + Ryzen 9 5950X + 64GB DDR4 (used market)

Memory: 24 GB VRAM + 64 GB system RAM
Runner: llama.cpp / Ollama (CUDA)

Runs comfortably
120 models

Ranked by fit for reasoning use case + predicted speed. Click a row for VRAM breakdown.

#1Llama 3.1 Nemotron Nano 8B
8B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 11.0 GBHeadroom: 13.0 GB
126
tok/s
Estimated
Weights
4.90 GB
KV cache
4.00 GB
Activations
0.25 GB
Runtime
1.80 GB
#2RefinedNeuro RN TR R1
8B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 11.0 GBHeadroom: 13.0 GB
ollama run RefinedNeuro/RN_TR_R1:latest
126
tok/s
Estimated
Weights
4.90 GB
KV cache
4.00 GB
Activations
0.25 GB
Runtime
1.80 GB
#3DeepSeek R1 Distill Llama 8B
8B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.7 GBHeadroom: 13.3 GB
126
tok/s
Estimated
Weights
4.70 GB
KV cache
4.00 GB
Activations
0.24 GB
Runtime
1.80 GB
Quant: Q4_K_MContext: 8,192VRAM: 11.6 GBHeadroom: 12.4 GB
112
tok/s
Estimated
Weights
5.00 GB
KV cache
4.50 GB
Activations
0.26 GB
Runtime
1.80 GB
#5DeepSeek R1 Distill Qwen 7B
7B
deepseek
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 13.8 GBHeadroom: 10.2 GB
ollama run deepseek-r1:7b
82
tok/s
Estimated
Weights
8.10 GB
KV cache
3.50 GB
Activations
0.41 GB
Runtime
1.80 GB
#6Phi-4 Reasoning 14B
14B
phi
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 17.6 GBHeadroom: 6.4 GB
ollama run phi4-reasoning:14b
72
tok/s
Estimated
Weights
8.40 GB
KV cache
7.00 GB
Activations
0.43 GB
Runtime
1.80 GB
#7DeepSeek R1 Distill Qwen 14B
14B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 17.6 GBHeadroom: 6.4 GB
ollama run deepseek-r1:14b
72
tok/s
Estimated
Weights
8.40 GB
KV cache
7.00 GB
Activations
0.43 GB
Runtime
1.80 GB
#8DeepSeek V3 Lite (16B MoE)
16B
deepseek
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 19.8 GBHeadroom: 4.2 GB
420
tok/s
Estimated
Weights
9.50 GB
KV cache
8.00 GB
Activations
0.48 GB
Runtime
1.80 GB
#9DeepSeek R1 Distill Mistral 24B
24B
deepseek
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 19.5 GBHeadroom: 4.5 GB
42
tok/s
Estimated
Weights
14.00 GB
KV cache
3.00 GB
Activations
0.70 GB
Runtime
1.80 GB
#10Phi-4 14B
14B
phi
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 17.6 GBHeadroom: 6.4 GB
ollama run phi4:14b
72
tok/s
Estimated
Weights
8.40 GB
KV cache
7.00 GB
Activations
0.43 GB
Runtime
1.80 GB
#11EXAONE Deep 7.8B
7.8B
other
Quant: Q4_K_MContext: 8,192VRAM: 10.2 GBHeadroom: 13.8 GB
129
tok/s
Estimated
Weights
4.30 GB
KV cache
3.90 GB
Activations
0.22 GB
Runtime
1.80 GB
#12RefinedNeuro RN TR R2
8B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 11.0 GBHeadroom: 13.0 GB
ollama run RefinedNeuro/RN_TR_R2:latest
126
tok/s
Estimated
Weights
4.90 GB
KV cache
4.00 GB
Activations
0.25 GB
Runtime
1.80 GB

Runs with tradeoffs
73 models

Tight VRAM, partial CPU offload, or context-limited.

Quant: Q4_K_MContext: 2,048VRAM: 23.6 GBHeadroom: 0.4 GB
  • Tight VRAM fit — only 0.4 GB headroom left for context growth
33
tok/s
Estimated
Weights
17.10 GB
KV cache
3.88 GB
Activations
0.86 GB
Runtime
1.80 GB
Sarvam M
24B
mistral
Commercial OK
Quant: Q4_K_MContext: 4,096VRAM: 21.7 GBHeadroom: 2.3 GB
  • Tight VRAM fit — only 2.3 GB headroom left for context growth
42
tok/s
Estimated
Weights
13.20 GB
KV cache
6.00 GB
Activations
0.66 GB
Runtime
1.80 GB
QwQ 32B Preview
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 37.8 GBHeadroom: 24.6 GB
  • Partial CPU offload: ~36% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
ollama run qwq:32b
2
tok/s
Estimated
Weights
19.00 GB
KV cache
16.00 GB
Activations
0.96 GB
Runtime
1.80 GB
Quant: Q4_K_MContext: 8,192VRAM: 36.3 GBHeadroom: 26.1 GB
  • Partial CPU offload: ~34% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
2
tok/s
Estimated
Weights
17.60 GB
KV cache
16.00 GB
Activations
0.89 GB
Runtime
1.80 GB
Qwen3 Swallow 32B RL v0.2
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 36.3 GBHeadroom: 26.1 GB
  • Partial CPU offload: ~34% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
2
tok/s
Estimated
Weights
17.60 GB
KV cache
16.00 GB
Activations
0.89 GB
Runtime
1.80 GB
DeepSeek R1 Distill Qwen 3 32B
32B
deepseek
Commercial OK
Quant: AWQ-INT4Context: 8,192VRAM: 37.8 GBHeadroom: 24.6 GB
  • Partial CPU offload: ~36% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
1
tok/s
Estimated
Weights
19.00 GB
KV cache
16.00 GB
Activations
0.96 GB
Runtime
1.80 GB
Magistral 32B
32B
mistral
Quant: AWQ-INT4Context: 8,192VRAM: 37.8 GBHeadroom: 24.6 GB
  • Partial CPU offload: ~36% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
1
tok/s
Estimated
Weights
19.00 GB
KV cache
16.00 GB
Activations
0.96 GB
Runtime
1.80 GB
Qwen 3 Coder 32B
32B
qwen
Commercial OK
Quant: AWQ-INT4Context: 8,192VRAM: 37.8 GBHeadroom: 24.6 GB
  • Partial CPU offload: ~36% of layers run on CPU
  • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
1
tok/s
Estimated
Weights
19.00 GB
KV cache
16.00 GB
Activations
0.96 GB
Runtime
1.80 GB

What if you upgraded?

Hypothetical scenarios. We re-ran the compatibility engine for each.

+32 GB system RAM

~$80–150

Doubles your CPU-offload working set. Helps when models don't quite fit in VRAM.

Unlocks: 64 new comfortable, 79 new tradeoff

  • Qwen 3 0.6B
  • Llama 3.2 3B Instruct
  • Qwen 3 1.7B
  • Gemma 3 270M

Upgrade to NVIDIA RTX PRO 4500 Blackwell

see current pricing

32 GB VRAM (vs your 24 GB) plus a bandwidth jump from ~936 GB/s to ~896 GB/s.

Unlocks: 103 new comfortable

  • Qwen 3 0.6B
  • Qwen 3 30B-A3B
  • Qwen 2.5 Coder 32B Instruct
  • Qwen3.6 27B

Add a second NVIDIA GeForce RTX 3090

~$899

Tensor parallelism splits the model across both cards, effectively doubling VRAM. Bandwidth doesn't double — runs ~1.5× the single-card speed in practice.

Unlocks: 116 new comfortable

  • Qwen 3 0.6B
  • Qwen 3 30B-A3B
  • Qwen 2.5 Coder 32B Instruct
  • Qwen3.6 27B

Some links above are affiliate links. We may earn a commission at no extra cost to you. How we make money.

Won't run
top 5 popular models

Need more memory than you have. Shown for orientation.

DeepSeek V4 Pro (1.6T MoE)
1600B
deepseek
Commercial OK

Even with CPU offload, needs more memory than your VRAM (24 GB) + 60% of system RAM (38 GB) combined.

Qwen 3.5 235B-A17B (MoE)
397B
qwen
Commercial OK

Even with CPU offload, needs more memory than your VRAM (24 GB) + 60% of system RAM (38 GB) combined.

Qwen 3 235B-A22B
235B
qwen
Commercial OK

Even with CPU offload, needs more memory than your VRAM (24 GB) + 60% of system RAM (38 GB) combined.

DeepSeek R1 (671B reasoning)
671B
deepseek
Commercial OK

Even with CPU offload, needs more memory than your VRAM (24 GB) + 60% of system RAM (38 GB) combined.

Llama 4 Scout
109B
llama
Commercial OK

Even with CPU offload, needs more memory than your VRAM (24 GB) + 60% of system RAM (38 GB) combined.

How to read these numbers

Measured here
Measured here - RunLocalAI ran this exact combo on owner hardware with public evidence.

Source-backed
Source-backed / community - a reproduced public source supports the speed, but it is not labeled as owner-measured.

Extrapolated
Extrapolated - predicted from a measured benchmark on similar-bandwidth hardware.

Estimated
Estimated - formula based on VRAM bandwidth and model architecture; not a benchmark row.

RunLocalAI Will-It-Run Framework →

Want a specific benchmark we don't have? Email Contact support and we'll prioritize it.