Used RTX 3090 vs RTX 5090 for local LLMs
The cheapest 24 GB card against the fastest consumer card, using only measured single-GPU results from sourced benchmarks.
| RTX 3090 (used) | RTX 5090 | |
|---|---|---|
| VRAM | 24 GB | 32 GB |
| Memory bandwidth | 936 GB/s | 1,792 GB/s |
| Price | ≈€1,248 used (eBay sold median, 237 sales, typical €1,136–€1,334; PCPrice.Watch, 2026-10-24) | ≈€6,199 new (Asus GeForce RTX 5090 (lowest variant), billiger.de, 2026-10-16) |
Measured generation speed, same model on each card (tok/s)
| Model | RTX 3090 | RTX 5090 | 5090 vs 3090 | 3090 tok/s per €100 | 5090 tok/s per €100 |
|---|---|---|---|---|---|
| Qwen3.8 27B | 42.9 (source) † | 75.2 (source) † | +75% | 3.4 | 1.2 |
| Qwen3.6 35B-A3B | 146-153 (source) † | 258 (source) † | +77% | 11.7 | 4.2 |
| Qwen3.6 27B | 42.7 (source) † | 75.7 (source) † | +77% | 3.4 | 1.2 |
| gpt-oss 20B | 207.2 (source) † | 324.2 (source) † | +56% | 16.6 | 5.2 |
Best plain-decode result per card from single-GPU builds (no speculative decoding / MTP, model fully in VRAM). Quants and runtimes can differ between sources, so treat differences as indicative and open each source.
Memory: 24 GB vs 32 GB
Weights at Q4_K_M (params × 4.89 bits ÷ 8) that fit in 24 GB: Qwen3.8 27B, Qwen3.6 35B-A3B, Qwen3.6 27B, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, gpt-oss 20B, Qwen3.5 9B, LFM2.5 24B-A2B, Nemotron 3.5 Lightning 30B-A3B, Ornith 1.5 35B-A3B.
Only in 32 GB: Flint 40B-A2.5B.
Only measured on the RTX 5090 so far
- Gemma 4 31B: 70.0 (source) † tok/s
- Gemma 4 26B-A4B: 235 (source) † tok/s
- Gemma 4 12B: 142 (source) † tok/s
- Qwen3.5 9B: 208 (source) † tok/s
Only measured on the RTX 3090 so far
- GLM 4.7 Flash: 134.6 (source) † tok/s
Big MoEs with RAM offload on a 5090: Ornith 1.5 122B, GLM-5.3 Flash and DeepSeek V4.1 Flash on one RTX 5090 + 128 GB RAM.
FAQ
Is a used RTX 3090 or an RTX 5090 better for local LLMs?
The RTX 5090 is faster: on the 4 current models measured on both cards in RigBase it generated 56–77% more tokens per second (median 76%), and it has 32 GB instead of 24 GB. The used RTX 3090 is far cheaper: about €1,248 (eBay sold median) vs about €6,199 new for the cheapest RTX 5090 in Germany, so it gave 2.8–3.2× more tokens/sec per euro on every model measured on both.
How much faster is the RTX 5090 than the RTX 3090 for LLM inference?
Median 76% faster in token generation across 4 models measured on both (Qwen3.8 27B, Qwen3.6 35B-A3B, Qwen3.6 27B, gpt-oss 20B). Memory bandwidth is 1,792 vs 936 GB/s (per llamabench.ai), which is what limits single-user generation speed.
What does a used RTX 3090 cost in Europe?
€1,248 median, typically €1,136–€1,334, from 237 completed eBay sales in 7 markets over 90 days (PCPrice.Watch, 2026-10-24). Other trackers differ, so check current sold listings before you buy.
What can 32 GB run that 24 GB can't?
At Q4_K_M, weights between 24 and 32 GB: Flint 40B-A2.5B (a smaller quant such as Q3_K_M fits 24 GB). Bigger mixture-of-experts models run on either card only with experts offloaded to system RAM.
Related: used RTX 3090 buying guide · RTX 3090 vs 4090 · RTX 4090 vs 5090 · best GPU under €1,000 · VRAM calculator
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.