BuildsGPUsModelsGuidesMethodologyNewsletter

RTX 3090 vs RTX 5090 for local LLMs

Same question every week: is the 5090 worth it over a used 3090?

SpecGeForce RTX 3090GeForce RTX 5090
Memory24 GB GDDR6X32 GB GDDR7
Bandwidth936.2 GB/s1790 GB/s
Low pricelow $1,889, avg $2,404 (used) source, 2026-10-15low $6,500, avg $8,757 (new) source, 2026-10-15
Qwen2.5-Coder-32B-Instruct estimate~36 tok/s~69 tok/s
$ per tok/s (Qwen2.5-Coder-32B-Instruct)$52$94
Models that fit22 of 2722 of 27
Largest that fitsQwen2.5-Coder-32B-InstructQwen2.5-Coder-32B-Instruct

Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.

What each one runs, model by model

Model (Q4_K_M)FileGeForce RTX 3090 estimateGeForce RTX 5090 estimate
Qwen2.5-Coder-32B-Instruct19.85 GB~36 tok/s~69 tok/s
DeepSeek-R1-Distill-Qwen-32B19.85 GB~36 tok/s~69 tok/s
Qwen3-32B19.76 GB~36 tok/s~69 tok/s
Qwen3-30B-A3B18.63 GB~353 tok/s~676 tok/s
gemma-3-27b-it16.55 GB~43 tok/s~82 tok/s
Mistral-Small-3.2-24B-Instruct-250614.33 GB~50 tok/s~95 tok/s
Devstral-Small-250514.33 GB~50 tok/s~95 tok/s
gpt-oss-20b11.67 GB~354 tok/s~677 tok/s
Qwen3-14B9 GB~79 tok/s~151 tok/s
Qwen2.5-Coder-14B-Instruct8.99 GB~79 tok/s~151 tok/s
DeepSeek-R1-Distill-Qwen-14B8.99 GB~79 tok/s~151 tok/s
Mistral-Nemo-Instruct-24077.48 GB~95 tok/s~182 tok/s
gemma-3-12b-it7.3 GB~97 tok/s~186 tok/s
NVIDIA-Nemotron-Nano-9B-v26.53 GB~109 tok/s~208 tok/s
Qwen3-8B5.03 GB~141 tok/s~270 tok/s
Llama-3.1-8B-Instruct4.92 GB~145 tok/s~277 tok/s
Qwen2.5-Coder-7B-Instruct4.68 GB~152 tok/s~291 tok/s
Mistral-7B-Instruct-v0.34.37 GB~163 tok/s~311 tok/s
Qwen3-4B2.5 GB~285 tok/s~544 tok/s
gemma-3-4b-it2.49 GB~286 tok/s~546 tok/s
Llama-3.2-3B-Instruct2.02 GB~352 tok/s~673 tok/s
Qwen3-0.6B0.48 GB~1482 tok/s~2834 tok/s

File sizes are Q4_K_M GGUFs from Hugging Face (linked on each model page), accessed 2026-10-15. A model fits when its file plus 1.5 GB of headroom is at most the card's memory.

Speed calculator estimate

Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works

Questions people ask

Which is faster on Qwen2.5-Coder-32B-Instruct?

On Qwen2.5-Coder-32B-Instruct (19.85 GB at Q4_K_M), the largest model that fits all of them: GeForce RTX 3090 about 36 tok/s, GeForce RTX 5090 about 69 tok/s. Speed tracks memory bandwidth: GeForce RTX 3090 936.2 GB/s, GeForce RTX 5090 1790 GB/s.

Which is cheaper per token per second?

Lowest tracked price divided by the estimated speed on Qwen2.5-Coder-32B-Instruct: GeForce RTX 3090 $52 per tok/s, GeForce RTX 5090 $94 per tok/s. Those prices are GeForce RTX 3090 $1,889 (used), GeForce RTX 5090 $6,500 (new), dated snapshots that move daily.

What fits on each of them?

GeForce RTX 3090 (24 GB): 22 of our 27 sized models, largest Qwen2.5-Coder-32B-Instruct at 19.85 GB; GeForce RTX 5090 (32 GB): 22 of our 27 sized models, largest Qwen2.5-Coder-32B-Instruct at 19.85 GB.

Is the extra memory worth it?

Not for fit: no model in our data fits the 32 GB GeForce RTX 5090 but not the 24 GB GeForce RTX 3090 — the next size up, DeepSeek-R1-Distill-Llama-70B at 42.52 GB, fits neither. The difference is bandwidth (GeForce RTX 3090 936.2 GB/s, GeForce RTX 5090 1790 GB/s), which on Qwen2.5-Coder-32B-Instruct works out to GeForce RTX 3090 about 36 tok/s, GeForce RTX 5090 about 69 tok/s.

Related