Under $300, one rule overrides everything else: VRAM is all that matters. Speed, architecture, and brand take a back seat to raw gigabytes. After the 2026 repricing there is exactly one card left inside this budget — a used RTX 3060 12GB at ~$250 — so the real question is not which card to buy but what its 12GB will and will not load.
NVIDIA GeForce RTX 3060 12GB
12GB GDDR612GB VRAM at ~$250 used. Runs all 7B models and 13B/14B at Q4_K_M. The only card still under $300.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What 12GB loads, model by model
This is the whole decision at this budget. VRAM figures are at Q4_K_M, Ollama’s default, with a couple of gigabytes left for KV cache — the same numbers used in our Ollama GPU requirements guide.
| Model | VRAM at Q4_K_M | Fits in 12GB? |
|---|---|---|
| Mistral 7B | ~5GB | Yes, with room for long context |
| Llama 3.1 8B | ~5.5GB | Yes, comfortably |
| Llama 2 13B | ~8.5GB | Yes |
| Qwen 2.5 14B | ~9GB | Yes |
| Llama 3.1 8B at Q8_0 | ~9GB | Yes, tight |
| Gemma 4 26B-A4B (MoE) | ~18GB | No — and 16GB does not fix it |
| Qwen 3.6 (35B-A3B MoE) | ~24GB | No — this is a 24GB-card model |
| Llama 3.1 8B at FP16 | ~16GB | No |
| CodeLlama 34B | ~20GB | No |
The two bold rows are the ones people get wrong, and it is worth being blunt about why. A Mixture-of-Experts model activates only a fraction of its weights per token — 3B of Qwen 3.6’s 35B, 4B of Gemma 4’s 26B — so it is widely assumed to need only that fraction of VRAM. It does not. Routing happens at inference time, so every expert has to be resident, and the sizes Ollama ships bear that out: 18GB for Gemma 4’s 26B-A4B and 24GB for Qwen 3.6. MoE buys speed, not capacity. At this budget those models are not one gigabyte away; they are a different class of card away.
Everything up to a 14B dense model, though, runs fine on 12GB — which is still most of what people actually use locally.
Under $300 GPU comparison
| GPU | VRAM | Bandwidth | Tok/s (7B Q4) | Price | Status |
|---|---|---|---|---|---|
| RTX 3060 12GB | 12GB | 360 GB/s | ~20 tok/s | ~$250 | Used — the only one under $300 |
| Intel Arc B580 | 12GB | 456 GB/s | ~18 tok/s* | ~$310 | New — just over |
| RTX 4060 Ti 16GB | 16GB | 288 GB/s | ~25 tok/s | ~$425 | New — 14B at Q8, not the MoE tier |
| RX 7800 XT | 16GB | 624 GB/s | ~28 tok/s | ~$450 | New — fastest 16GB here, ROCm |
| RTX 4060 | 8GB | 272 GB/s | ~22 tok/s | ~$479 | New — 8GB, and dearer than the 16GB cards |
*Arc B580 Linux only for best LLM support — Windows support via IPEX-LLM is functional but less mature.
Only the first row is under $300. The RTX 4060 is worth a second look: at ~$479 for 8GB it now costs more than two different 16GB cards, which caps it at 7B models while charging flagship-adjacent money. It is not a recommendation at any budget on this page.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
#1: RTX 3060 12GB (used) — best VRAM per dollar
The RTX 3060 12GB is the community’s go-to budget LLM card for a reason that has not changed in two years: no new card under $300 gives you 12GB.
What 12GB unlocks that 8GB cannot:
- Llama 3.1 8B at Q6_K (~7GB) — near-lossless quality on the best 8B model
- Mistral 7B at Q8 (~8GB) — full quality
- Qwen 3 14B at Q4_K_M (9.3GB) — runs with headroom
- DeepSeek Coder V2 Lite (16B) at Q3_K_M (~9GB) — workable for code tasks
- Llama 2 13B at Q4_K_M (~8GB) — comfortable
The 360 GB/s bandwidth is notably high for this tier — actually faster than the newer RTX 4060 (272 GB/s). That translates to better token throughput on memory-bound inference tasks.
Where to buy: eBay, used GPU markets, PC part swap communities. Look for RTX 3060 12GB specifically — not the RTX 3060 Ti (which only has 8GB) and not the desktop 12GB that was occasionally cut to 8GB in some OEM builds. Verify VRAM before buying.
The nearest new card: Intel Arc B580 at ~$310
The Arc B580 sat at ~$250 when this guide was written and now sells around $310, which puts it just outside the title. It remains the cheapest new route to 12GB:
- 12GB GDDR6 — matches the RTX 3060’s VRAM on new silicon
- 456 GB/s bandwidth — nominally faster than the RTX 3060’s 360 GB/s
- Intel XMX AI acceleration — helps with some inference paths
- Full warranty as new hardware
The trade-off: Intel’s LLM support ecosystem is not NVIDIA’s. Ollama works natively on Windows and Linux now, but community benchmarks, guides, and troubleshooting resources are sparse compared to NVIDIA. If you hit a problem, you may be on your own. That gap also shows up in throughput — despite the bandwidth advantage on paper, it lands around 18 tok/s on a 7B against the 3060’s 20, because the software stack is doing less with the hardware.
For Linux users comfortable with some configuration, the B580 is a legitimate alternative to a used card. For Windows users who want a smooth Ollama setup out of the box, the RTX 3060 12GB used is safer. For a full breakdown of what the B580 handles in practice, see Intel Arc B580 for LLM inference.
Check Intel Arc B580 on Amazon→Buy on Shopee SG→Where 12GB stops paying: the 16GB step
The 16GB step is real, but it is smaller than it is usually sold as. An RTX 4060 Ti 16GB at ~$425 buys you:
- a 14B at Q8 (~16GB) instead of Q4, and an 8B at FP16 — precision, not size
- headroom for long context on a 13-14B model, where 12GB starts swapping
- ~25 tok/s on a 7B, ~18 on a 13B — modestly faster than the 3060
- 288 GB/s is its weak point; the RX 7800 XT at ~$450 doubles that at 624 GB/s if ROCm suits you
What it does not buy is the 2026 MoE tier. Gemma 4’s 26B-A4B fits only in its 16GB QAT build with nothing left for context, and Qwen 3.6 does not fit at all. The first card that runs both properly is a used RTX 3090 24GB at ~$820 — $570 past this guide’s ceiling, not $175. Our under $500 guide covers the middle bracket properly.
NVIDIA GeForce RTX 4060 Ti 16GB
16GB GDDR6The honest 16GB case: a 14B at Q8 instead of Q4, an 8B at FP16, and real context headroom. Not the MoE tier — that starts at 24GB.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Which GPU should YOU buy under $300?
If $300 is a hard ceiling: RTX 3060 12GB used (~$250). It is the only card inside the line, and it is battle-tested with every major LLM tool. Buy from a reputable seller and verify the VRAM.
If you will not buy used: Intel Arc B580 (~$310), which puts you $10 over. There is no new card under $300 in 2026.
If the MoE models matter to you: neither this budget nor the next one gets you there. Gemma 4 26B-A4B needs ~18GB and Qwen 3.6 ~24GB, so the entry ticket is a used RTX 3090 24GB at ~$820. Buying a 16GB card for those two models is the most common way to waste $425 in 2026.
Throughput on the cards in and near this budget
Figures are modelled from memory bandwidth rather than measured — see our methodology for how and why.
| Model | RTX 3060 12GB (~$250) | Arc B580 12GB (~$310) | RTX 4060 Ti 16GB (~$425) |
|---|---|---|---|
| 7B at Q4 | ~20 tok/s | ~18 tok/s | ~25 tok/s |
| 13B at Q4 | ~12 tok/s | ~10 tok/s | ~18 tok/s |
| 14B at Q8 (~16GB) | Won’t fit | Won’t fit | ~14 tok/s |
| Qwen 3.6 35B MoE (~24GB) | Won’t fit | Won’t fit | Won’t fit |
| 34B dense at Q4 | Won’t fit | Won’t fit | Won’t fit |
The speed difference across this whole range is small. What separates the rows is capacity — and the honest reading of the bottom two rows is that the 2026 MoE models are outside this budget entirely, not one card away from it. What $425 buys over $250 is precision on models you can already run, not a new tier of them.
Common mistakes to avoid
- Buying an 8GB card at this budget. The RTX 4060 is the trap: 8GB caps you at 7B models, and at ~$479 it now costs more than two different 16GB cards. Speed on a model you cannot load is worth nothing.
- Confusing the RTX 3060 Ti with the RTX 3060. The 3060 Ti has 8GB — not 12GB. Always verify the exact model before buying used.
- Reading a MoE model’s active parameter count as its VRAM requirement. “35B-A3B” activates 3B parameters per token but stores all 35B, and Ollama ships it as a 24GB download. This is the single most expensive misconception at this budget, because it makes a 16GB card look like it clears a bar it does not.
- Working from a 2025 price list. The Arc B580 at $250 and the RTX 4060 at $280 were both true a year ago. Check current listings before trusting any budget guide, this one included.
Final verdict
| Goal | GPU | Price |
|---|---|---|
| Hard $300 ceiling | RTX 3060 12GB (used) | ~$250 |
| New card, nearest to budget | Intel Arc B580 | ~$310 |
| Reach the 2026 MoE tier | RTX 3090 24GB (used) | ~$820 |
Under $300 the used market is not merely the better option, it is the only one. A 12GB RTX 3060 runs everything up to a 14B dense model, which covers most of what people actually do locally — and the 2026 MoE models it cannot reach are further away than their names suggest, on the far side of a 24GB card rather than a 16GB one.
For more on budget GPU options, see the best budget GPU for local LLM guide and our under $500 GPU guide for the next tier up. If you are weighing used GPU options specifically, the best used GPU for LLM guide covers the landscape.