What can NVIDIA GeForce RTX 4090 run for vision?

Build: RTX 4090 + Ryzen 9 7950X + 64GB DDR5

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

Runs comfortably
18 models

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

#1Gemma 4 E4B (Effective 4B)
4B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 8.4 GBHeadroom: 15.6 GBTTFT: fast
ollama run gemma4:e4b
154
tok/s
Estimated
Weights
4.40 GB
KV cache
2.00 GB
Activations
0.23 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~124 ms (fast)
Model details →
#2Gemma 3 4B
4B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 8.4 GBHeadroom: 15.6 GBTTFT: fast
ollama run gemma3:4b
154
tok/s
Estimated
Weights
4.40 GB
KV cache
2.00 GB
Activations
0.23 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~124 ms (fast)
Model details →
#3Gemma 4 E2B (Effective 2B)
2B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 5.1 GBHeadroom: 18.9 GBTTFT: instant
ollama run gemma4:e2b
308
tok/s
Estimated
Weights
2.20 GB
KV cache
1.00 GB
Activations
0.12 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~62 ms (instant)
Model details →
#4Phi-3.5 Vision
4.2B
phi
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 6.5 GBHeadroom: 17.5 GBTTFT: fast
258
tok/s
Estimated
Weights
2.50 GB
KV cache
2.10 GB
Activations
0.13 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~130 ms (fast)
Model details →
#5MiniCPM-V 2.6 8B
8B
minicpm
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 11.1 GBHeadroom: 12.9 GBTTFT: fast
136
tok/s
Estimated
Weights
5.00 GB
KV cache
4.00 GB
Activations
0.26 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~248 ms (fast)
Model details →
#6Qwen 2.5-VL 7B
7B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 9.9 GBHeadroom: 14.1 GBTTFT: fast
155
tok/s
Estimated
Weights
4.40 GB
KV cache
3.50 GB
Activations
0.23 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#7Moondream 2
1.9B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 3.3 GBHeadroom: 20.7 GBTTFT: instant
571
tok/s
Estimated
Weights
1.20 GB
KV cache
0.24 GB
Activations
0.06 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~59 ms (instant)
Model details →
#8Qwen 2-VL 7B
7B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.1 GBHeadroom: 13.9 GBTTFT: fast
155
tok/s
Estimated
Weights
4.60 GB
KV cache
3.50 GB
Activations
0.24 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#9LLaVA 1.6 Mistral 7B
7B
other
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.0 GBHeadroom: 14.0 GBTTFT: fast
155
tok/s
Estimated
Weights
4.50 GB
KV cache
3.50 GB
Activations
0.23 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#10LLaVA-OneVision 7B
7B
other
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 10.0 GBHeadroom: 14.0 GBTTFT: fast
155
tok/s
Estimated
Weights
4.50 GB
KV cache
3.50 GB
Activations
0.23 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~217 ms (fast)
Model details →
#11Molmo 7B-D
8B
other
Commercial OK
Quant: Q4_K_MContext: 4,096VRAM: 9.3 GBHeadroom: 14.7 GBTTFT: fast
136
tok/s
Estimated
Weights
5.20 GB
KV cache
2.00 GB
Activations
0.26 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~248 ms (fast)
Model details →
#12Qwen 2.5-VL 3B
3B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 5.4 GBHeadroom: 18.6 GBTTFT: instant
362
tok/s
Estimated
Weights
2.00 GB
KV cache
1.50 GB
Activations
0.11 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~93 ms (instant)
Model details →

Runs with tradeoffs
10 models

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

Llama 3.2 11B Vision Instruct
11B
llama
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 20.4 GBHeadroom: 3.6 GBTTFT: fast
  • • Tight VRAM fit — only 3.6 GB headroom left for context growth
ollama run llama3.2-vision:11b
56
tok/s
Estimated
Weights
12.50 GB
KV cache
5.50 GB
Activations
0.63 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~341 ms (fast)
Model details →
InternVL 2.5 26B
26B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 21.9 GBHeadroom: 2.1 GBTTFT: noticeable
  • • Tight VRAM fit — only 2.1 GB headroom left for context growth
42
tok/s
Estimated
Weights
16.00 GB
KV cache
3.25 GB
Activations
0.80 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~806 ms (noticeable)
Model details →
Gemma 3 27B
27B
gemma
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 22.0 GBHeadroom: 2.0 GBTTFT: noticeable
  • • Tight VRAM fit — only 2.0 GB headroom left for context growth
ollama run gemma3:27b
40
tok/s
Estimated
Weights
16.00 GB
KV cache
3.38 GB
Activations
0.80 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~837 ms (noticeable)
Model details →
Qwen3.8 27B
27B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 22.5 GBHeadroom: 1.5 GBTTFT: noticeable
  • • Tight VRAM fit — only 1.5 GB headroom left for context growth
ollama run qwen3.8:27b
40
tok/s
Estimated
Weights
16.50 GB
KV cache
3.38 GB
Activations
0.83 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~837 ms (noticeable)
Model details →
MedGemma 27B
27B
gemma
Quant: Q4_K_MContext: 2,048VRAM: 22.0 GBHeadroom: 2.0 GBTTFT: noticeable
  • • Tight VRAM fit — only 2.0 GB headroom left for context growth
40
tok/s
Estimated
Weights
16.00 GB
KV cache
3.38 GB
Activations
0.80 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~837 ms (noticeable)
Model details →
PaliGemma 2 10B
10B
gemma
Commercial OK
Quant: BF16Context: 8,192VRAM: 27.8 GBHeadroom: 34.6 GBTTFT: fast
  • • Partial CPU offload: ~14% of layers run on CPU
  • • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
10
tok/s
Estimated
Weights
20.00 GB
KV cache
5.00 GB
Activations
1.01 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~310 ms (fast)
Model details →
Gemma 4 31B Dense
31B
gemma
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 52.0 GBHeadroom: 10.4 GBTTFT: noticeable
  • • Partial CPU offload: ~54% of layers run on CPU
  • • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
ollama run gemma4:31b
2
tok/s
Estimated
Weights
33.00 GB
KV cache
15.50 GB
Activations
1.66 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~961 ms (noticeable)
Model details →
Molmo 72B
72B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 53.9 GBHeadroom: 8.5 GBTTFT: slow
  • • Partial CPU offload: ~55% of layers run on CPU
  • • CPU is the bottleneck — upgrading RAM bandwidth helps more than VRAM here
1
tok/s
Estimated
Weights
41.00 GB
KV cache
9.00 GB
Activations
2.05 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~2231 ms (slow)
Model details →

What if you upgraded?

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

+32 GB system RAM

Check the current price

Adds 32 GB to your CPU-offload working set. Helps when models don't quite fit in VRAM.

Unlocks: 95 new comfortable, 39 new tradeoff

  • • Qwen 3 0.6B
  • • Llama 3.1 8B Instruct
  • • Qwen 3 8B
  • • Qwen 3 14B
Shop this upgrade↗

Upgrade to NVIDIA GeForce RTX 5090

Launch MSRP $1,999 (2025). Check the current price.

32 GB VRAM (vs your 24 GB) plus a bandwidth jump from ~1008 GB/s to ~1792 GB/s.

Unlocks: 110 new comfortable

  • • Qwen 3 0.6B
  • • Llama 3.1 8B Instruct
  • • Qwen 3 30B-A3B
  • • Qwen 2.5 Coder 32B Instruct
Shop this upgrade↗

Add a second NVIDIA GeForce RTX 4090

Launch MSRP $1,599 (2022). Check the current price.

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: 118 new comfortable

  • • Qwen 3 0.6B
  • • Llama 3.1 8B Instruct
  • • Qwen 3 30B-A3B
  • • Qwen 2.5 Coder 32B Instruct
Shop this upgrade↗

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 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.

—
DeepSeek V4 Flash (284B MoE)
284B
deepseek
Commercial OK

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

—
Cloud GPU rental

Not enough VRAM for these models? Rent a GPU by the hour

The models above need more memory than 24 GB VRAM + 64 GB system RAM. You can rent a GPU with enough VRAM by the hour to run one of them, or to test it before you buy hardware. For daily use, owning a card usually costs less over time.

RunPod

billed per GPU-hour

Hourly GPU pods, in community-cloud and secure-cloud tiers.

The two tiers run on different hosts and terms. Check which tier a pod is on before a long job.

Rent on RunPod

Vast.ai

billed per GPU-hour

Marketplace of host-listed GPUs; price and reliability vary by host.

Host quality varies. Filter for verified hosts and check the host's reliability before a long run.

Rent on Vast.ai

Referral links: RunLocalAI may earn a commission or account credits if you sign up and rent, at no extra cost to you. Hourly rates and which GPUs are listed change all the time, so check the provider's page before you start. Editorial opinions are independent of the referral relationship; see how we make money.

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.