Best GPU for local LLMs under €1,000
Which graphics card should you buy for running local LLMs on a budget under €1,000? Every card below has a current German retail price and, where available, a measured tokens/sec result from a sourced benchmark.
Cards under (and just over) €1,000
| GPU | VRAM | Lowest new price DE | Measured 2026-model results (single card, plain decode) | |
|---|---|---|---|---|
| RTX 3060 | 12 GB | ≈€429 | Qwen3.8 27B: 5.2-6.7 tok/s (1.2 per €100, source) † Gemma 4 12B: 38.5 tok/s (9.0 per €100, source) gpt-oss 20B: 83.96 tok/s (19.6 per €100, source) † LFM2.5 24B-A2B: 125 tok/s (29.1 per €100, source) † | under budget |
| RTX 5060 Ti 16GB | 16 GB | ≈€739 | gpt-oss 20B: 111.51 tok/s (15.1 per €100, source) | under budget |
| Radeon RX 7900 XT | 20 GB | ≈€869 | no 2026-model result yet | under budget |
| RTX A4000 | 16 GB | ≈€1,050 | gpt-oss 20B: 93.55 tok/s (8.9 per €100, source) Muse Glimmer 30B: 21.42 tok/s (2.0 per €100, source) | just over |
| RTX 5070 Ti | 16 GB | ≈€1,188 | Gemma 4 12B: 87.6 tok/s (7.4 per €100, source) gpt-oss 20B: 189.45 tok/s (16.0 per €100, source) | just over |
Prices: billiger.de lowest new price incl. VAT, 2026-10-16. Results: best plain-decode single-GPU result per model with the model fully in VRAM; runtimes and quants differ by source. † The model's Q4_K_M / native weights are larger than this card's VRAM, so the source ran part of it from system RAM or used a smaller quant; it is not counted in the short answer.
What fits in 12, 16 and 20 GB (Q4_K_M weights)
- 12 GB: Gemma 4 12B, Qwen3.5 9B
- 16 GB adds: Gemma 4 26B-A4B, gpt-oss 20B, LFM2.5 24B-A2B
- 20 GB adds: Qwen3.8 27B, Qwen3.6 27B, Gemma 4 31B, Nemotron 3.5 Lightning 30B-A3B
Weights = params × 4.89 bits ÷ 8 (llama.cpp Q4_K_M), before context. Per-model math is on each model page and in the VRAM calculator.
The used option: RTX 3090 (24 GB)
Median €1,248 used (typical €1,136–€1,334; PCPrice.Watch eBay sold data, 2026-10-24), so slightly over budget, but it is the cheapest way to 24 GB. See used RTX 3090 vs RTX 5090 and the used RTX 3090 buying guide.
FAQ
What is the best GPU for local LLMs under €1,000?
From measured results in RigBase, the RTX 5060 Ti 16GB (≈€739, 16 GB) is both the fastest new card under €1,000 with gpt-oss 20B fully in VRAM (111.51 tok/s) and the best tok/s per euro on it (15.1 per €100). VRAM decides which models fit at all, so check the memory table before choosing on speed.
Is 12 GB or 16 GB enough for local LLMs?
At Q4_K_M, 2026 models whose weights fit in 12 GB: Gemma 4 12B, Qwen3.5 9B. Up to 16 GB adds: Gemma 4 26B-A4B, gpt-oss 20B, LFM2.5 24B-A2B. Leave extra room for context (KV cache).
Can I get a used RTX 3090 for under €1,000?
Not at the median right now: used RTX 3090s sold for a median €1,248 on eBay (typical €1,136–€1,334, 237 sales, PCPrice.Watch, 2026-10-24). It is the cheapest 24 GB card if you can stretch the budget.
Where do these prices come from?
Lowest new prices on billiger.de (Germany, incl. VAT) as of 2026-10-16; used RTX 3090 from eBay sold listings via PCPrice.Watch. Prices change daily, so check the listing before buying.
Related: complete AI rigs under €1,000 · best value GPU per euro · best GPU 2026 (all budgets)
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.