What can NVIDIA GeForce RTX 5090 run for agents?

Build: NVIDIA GeForce RTX 5090 + — + 32 GB RAM (windows)

Memory: 32 GB VRAM + 32 GB system RAM
Runner: llama.cpp / Ollama (CUDA)
AnyChatCodingAgentsReasoningVisionLong contextCreative

Runs comfortably
82 models

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

#1Hermes 3 Llama 3.1 8B
8B
hermes
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 14.7 GBHeadroom: 17.3 GBTTFT: fast
ollama run hermes3:8b
137
tok/s
Estimated
Weights
8.50 GB
KV cache
4.00 GB
Activations
0.43 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~164 ms (fast)
Model details →
#2Qwen 2.5 14B Instruct
14B
qwen
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 25.3 GBHeadroom: 6.7 GBTTFT: fast
ollama run qwen2.5:14b
78
tok/s
Estimated
Weights
15.70 GB
KV cache
7.00 GB
Activations
0.79 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~287 ms (fast)
Model details →
#3Qwen 3 30B-A3B
30B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 24.5 GBHeadroom: 7.5 GBTTFT: noticeable
ollama run qwen3:30b
64
tok/s
Estimated
Weights
18.00 GB
KV cache
3.75 GB
Activations
0.90 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~614 ms (noticeable)
Model details →
#4Qwen3.8 27B
27B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 22.5 GBHeadroom: 9.5 GBTTFT: noticeable
ollama run qwen3.8:27b
71
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): ~553 ms (noticeable)
Model details →
#5Qwen 2.5 Coder 32B Instruct
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 23.9 GBHeadroom: 8.1 GBTTFT: noticeable
ollama run qwen2.5-coder:32b
60
tok/s
Estimated
Weights
19.00 GB
KV cache
2.15 GB
Activations
0.96 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~655 ms (noticeable)
Model details →
#6Qwen 3 32B
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 25.8 GBHeadroom: 6.2 GBTTFT: noticeable
ollama run qwen3:32b
60
tok/s
Estimated
Weights
19.00 GB
KV cache
4.00 GB
Activations
0.95 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~655 ms (noticeable)
Model details →
#7Qwen 2.5 32B Instruct
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 25.8 GBHeadroom: 6.2 GBTTFT: noticeable
ollama run qwen2.5:32b
60
tok/s
Estimated
Weights
19.00 GB
KV cache
4.00 GB
Activations
0.95 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~655 ms (noticeable)
Model details →
#8Qwen 2.5 7B Instruct
7B
qwen
Commercial OK
Quant: Q8_0Context: 8,192VRAM: 10.8 GBHeadroom: 21.2 GBTTFT: fast
ollama run qwen2.5:7b
157
tok/s
Estimated
Weights
8.10 GB
KV cache
0.47 GB
Activations
0.41 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~143 ms (fast)
Model details →
#9Llama 3.1 Nemotron Nano 8B
8B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 11.0 GBHeadroom: 21.0 GBTTFT: fast
241
tok/s
Estimated
Weights
4.90 GB
KV cache
4.00 GB
Activations
0.25 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~164 ms (fast)
Model details →
#10Mistral 7B Instruct v0.3
7B
mistral
Commercial OK
Quant: Q5_K_MContext: 8,192VRAM: 10.7 GBHeadroom: 21.3 GBTTFT: fast
ollama run mistral:7b
242
tok/s
Estimated
Weights
5.10 GB
KV cache
3.50 GB
Activations
0.26 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~143 ms (fast)
Model details →
#11Llama 3.1 8B Instruct
8B
llama
Commercial OK
Quant: FP16Context: 8,192VRAM: 19.8 GBHeadroom: 12.2 GBTTFT: fast
ollama run llama3.1:8b
73
tok/s
Estimated
Weights
16.10 GB
KV cache
1.07 GB
Activations
0.81 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~164 ms (fast)
Model details →
#12QwQ 32B Preview
32B
qwen
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 25.8 GBHeadroom: 6.2 GBTTFT: noticeable
ollama run qwq:32b
60
tok/s
Estimated
Weights
19.00 GB
KV cache
4.00 GB
Activations
0.95 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~655 ms (noticeable)
Model details →

Runs with tradeoffs
9 models

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

Dolphin 3.0 Mistral 24B
24B
dolphin
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 28.5 GBHeadroom: 3.5 GBTTFT: fast
  • • Tight VRAM fit — only 3.5 GB headroom left for context growth
ollama run dolphin-mistral:24b
80
tok/s
Estimated
Weights
14.00 GB
KV cache
12.00 GB
Activations
0.71 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~492 ms (fast)
Model details →
Mixtral 8x7B Instruct
47B
mixtral
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 37.1 GBHeadroom: 14.1 GBTTFT: noticeable
  • • Partial CPU offload: ~14% of layers run on CPU
ollama run mixtral:8x7b
41
tok/s
Estimated
Weights
28.00 GB
KV cache
5.88 GB
Activations
1.40 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~963 ms (noticeable)
Model details →
Mistral Small 3 24B
24B
mistral
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 28.5 GBHeadroom: 3.5 GBTTFT: fast
  • • Tight VRAM fit — only 3.5 GB headroom left for context growth
ollama run mistral-small:24b
80
tok/s
Estimated
Weights
14.00 GB
KV cache
12.00 GB
Activations
0.71 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~492 ms (fast)
Model details →
Llama 3.3 70B Instruct
70B
llama
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 46.5 GBHeadroom: 4.7 GBTTFT: noticeable
  • • Partial CPU offload: ~31% of layers run on CPU
ollama run llama3.3:70b
28
tok/s
Estimated
Weights
40.00 GB
KV cache
2.68 GB
Activations
2.01 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~1434 ms (noticeable)
Model details →
Jamba 1.5 Mini
52B
other
Commercial OK
Quant: Q4_K_MContext: 2,048VRAM: 39.8 GBHeadroom: 11.4 GBTTFT: fast
  • • Partial CPU offload: ~20% of layers run on CPU
161
tok/s
Estimated
Weights
30.00 GB
KV cache
6.50 GB
Activations
1.50 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~246 ms (fast)
Model details →
Command R 35B
35B
command-r
Quant: Q4_K_MContext: 2,048VRAM: 28.2 GBHeadroom: 3.8 GBTTFT: noticeable
  • • Tight VRAM fit — only 3.8 GB headroom left for context growth
ollama run command-r:35b
55
tok/s
Estimated
Weights
21.00 GB
KV cache
4.38 GB
Activations
1.05 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~717 ms (noticeable)
Model details →
Aya 23 35B
35B
other
Quant: Q4_K_MContext: 2,048VRAM: 28.2 GBHeadroom: 3.8 GBTTFT: noticeable
  • • Tight VRAM fit — only 3.8 GB headroom left for context growth
55
tok/s
Estimated
Weights
21.00 GB
KV cache
4.38 GB
Activations
1.05 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~717 ms (noticeable)
Model details →
Gemma 4 Turkish 26B (4B active)
26B
gemma
Commercial OK
Quant: Q4_K_MContext: 8,192VRAM: 29.8 GBHeadroom: 2.2 GBTTFT: noticeable
  • • Tight VRAM fit — only 2.2 GB headroom left for context growth
74
tok/s
Estimated
Weights
14.30 GB
KV cache
13.00 GB
Activations
0.72 GB
Runtime
1.80 GB
Time to first token (prefill, 512-token prompt): ~532 ms (noticeable)
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: 46 new comfortable, 22 new tradeoff

  • • Qwen 3 0.6B
  • • Llama 3.2 3B Instruct
  • • Qwen 3 1.7B
  • • Gemma 3 270M
Shop this upgrade↗

Upgrade to NVIDIA A100 40GB

Launch MSRP $11,000 (2020). Check the current price.

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

Unlocks: 52 new comfortable

  • • Qwen 3 0.6B
  • • Llama 3.2 3B Instruct
  • • Qwen 3 1.7B
  • • Gemma 3 270M
Shop this upgrade↗

Add a second NVIDIA GeForce RTX 5090

Launch MSRP $1,999 (2025). 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: 64 new comfortable

  • • Qwen 3 0.6B
  • • Llama 3.3 70B Instruct
  • • DeepSeek R1 Distill Llama 70B
  • • Llama 3.2 3B 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 (32 GB) + 60% of system RAM (19 GB) combined.

—
Qwen 3 235B-A22B
235B
qwen
Commercial OK

Even with CPU offload, needs more memory than your VRAM (32 GB) + 60% of system RAM (19 GB) combined.

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

Even with CPU offload, needs more memory than your VRAM (32 GB) + 60% of system RAM (19 GB) combined.

—
Llama 4 Scout
109B
llama
Commercial OK

Even with CPU offload, needs more memory than your VRAM (32 GB) + 60% of system RAM (19 GB) combined.

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

Even with CPU offload, needs more memory than your VRAM (32 GB) + 60% of system RAM (19 GB) combined.

—
Cloud GPU rental

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

The models above need more memory than 32 GB VRAM + 32 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.