You have just been handed a $1,500 hardware budget for a local LLM workstation. Until this year that budget bought an RTX 4090. The GDDR7 shortage moved it to roughly $2,200, so the tier now has a different shape: one used card holds 24GB, and everything new tops out at 16GB.
Quick answer: A used RTX 3090 (24GB, ~$820) is the best GPU for local LLM under $1,500 in 2026. 24GB is the line where 32-34B models load at Q4 without touching system RAM, and no new card in this budget crosses it.
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB at roughly $820 used - loads 32B and 34B models at Q4 with context headroom, which no 16GB card in this budget can do.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Who this is for
You have a budget between $1,000 and $1,500 for a GPU dedicated to local LLM inference. You want the best possible model quality and speed for that money, whether you are a developer, researcher, or power user running models daily.
GPU comparison under $1500
| GPU | VRAM | Bandwidth | 7B Q4 Speed | 14B Q4 Speed | 32B Q4 | Price |
|---|---|---|---|---|---|---|
| RTX 3090 (used) | 24GB | 936 GB/s | ~55 tok/s | ~32 tok/s | ~18 tok/s | ~$820 |
| RTX 5080 | 16GB | 960 GB/s | ~55 tok/s | ~32 tok/s | Won’t fit | ~$1,400 |
| RTX 5070 Ti | 16GB | 896 GB/s | ~45 tok/s | ~30 tok/s | Won’t fit | ~$1,050 |
| RTX 4090 (over budget) | 24GB | 1,008 GB/s | ~65 tok/s | ~38 tok/s | ~20 tok/s | ~$2,200 |
Read the 32B column first. It is the only one where the cards disagree about what is possible rather than how fast. The RTX 5080 costs $580 more than a used 3090 and still returns “won’t fit”, because 16GB cannot hold a 32B model at Q4 no matter how quick its memory is. Bandwidth sets your tok/s; capacity sets your menu.
What 24GB actually loads
These are the model sizes the used RTX 3090’s 24GB opens up, and the reason it beats faster 16GB cards in this budget:
- DeepSeek-R1 32B at Q4_K_M (~19GB) — fits with headroom for context
- Qwen 2.5 32B at Q4_K_M (~19GB) — comfortable
- CodeLlama 34B at Q4_K_M (~20GB) — fits
- Any 7B-14B model at FP16 — no quantization needed
- 7B models with 32K+ context — VRAM to spare
A 16GB card can be made to “run” a 32B model by offloading layers to system RAM, and this is where the tok/s column stops being the useful number. Offloaded layers move across PCIe on every token, and throughput typically falls to a small fraction of the fully-resident figure. The honest reading is not that a 16GB card is slower at 32B; it is that 32B is off the menu.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→Alternatives worth considering
RTX 5080 (~$1,400): The best new 16GB card, and faster than a 3090 on anything that fits in 16GB. It takes almost the whole budget and closes the 32B door. Choose it if you are confident you will stay in the 7B-14B range and want warranty coverage. For the head-to-head on LLM workloads, see RTX 5080 vs 4090 for LLM.
RTX 5070 Ti (~$1,050): Saves $350 against the 5080 for roughly 90% of its throughput, with the same 16GB ceiling. The sensible new card if the budget also has to cover RAM or an SSD.
RTX 4090 (~$2,200): Out of budget as of 2026. It is the fastest 24GB card here and would still be the pick at its old price, but that price is gone. Our under $2,000 guide covers what the next tier up actually reaches.
Which GPU should you buy under $1500?
Answer one question: will you ever run a 32B model? If yes, a used RTX 3090 ($820) is the only card in budget that can, and the decision is made. If you are certain you will stay at 7B-14B, the RTX 5080 ($1,400) is faster on those sizes and comes with a warranty, while the RTX 5070 Ti ($1,050) gives up little and leaves ~$450 for RAM, an SSD, or the PSU this build will need anyway.
Common mistakes to avoid
- Buying two mid-range GPUs instead of one 24GB card. Two RTX 4060 Ti 16GBs (~$850) reach 32GB on paper, but splitting a model across two cards through llama.cpp adds latency and setup work that a single 24GB card avoids entirely.
- Reading the tok/s column and ignoring the VRAM column. The RTX 5080 posts better numbers than a used 3090 on every size that fits in 16GB. That is the whole caveat: on 32B it posts nothing at all.
- Ignoring PSU requirements. A used RTX 3090 pulls up to 350W and the RTX 5080 up to 360W. Budget $100-150 for an 850W supply if yours is marginal.
- Waiting for the RTX 4090 to come back into range. It moved the other way through 2026, from roughly $1,600 to roughly $2,200, after production ended and GDDR7 supply went to AI accelerators.
Our recommendation
| Your need | Best GPU | Price |
|---|---|---|
| Running 32B models | RTX 3090 (used) | ~$820 |
| Best value 24GB | RTX 3090 (used) | ~$820 |
| Best new 16GB card | RTX 5080 | ~$1,400 |
| New card, budget left over | RTX 5070 Ti | ~$1,050 |
This tier used to be defined by how much of the budget to spend. In 2026 it is defined by capacity: 24GB exists only on the used market, and every new card here stops at 16GB. Pick the ceiling that matches the models you actually run.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→NVIDIA GeForce RTX 5080
16GB GDDR716GB of GDDR7 with a warranty at roughly $1,400 - the fastest option here for 7B-14B inference, with 32B out of reach.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
A $1,500 LLM budget used to buy the fastest 24GB card made. It now buys a two-generation-old one, and for deciding which models you can load, that trade still comes out ahead.
For tighter budgets, see our under $1,000 GPU guide. If you can stretch further, the under $2,000 GPU guide covers the RTX 5090’s 32GB tier. If you want to explore what each GPU can run through Ollama, check our Ollama GPU guide.