Best budget GPU for local LLMs in 2026: 6 GB, 8 GB or 12 GB, measured

Buy for the model size you want, because VRAM is the real limit

The card you want is the one that holds your model entirely in VRAM, and that is decided by gigabytes, not CUDA cores. On our test bench, a 7B/8B Q4 model ran fully on an 8 GB RTX 3060 Ti at 70+ tokens/s (bench notes). A 14B Q4 model needs about 9 GB to stay fully on GPU and lost more than a third of its speed when it spilled to CPU.

Key takeaway: 8 GB runs 7B/8B well. 14B needs about 9 GB, so a 12 GB card is the cheapest way to get it. In September 2026 the refurbished RTX 3060 12 GB is the value pick.

Ollama publishes download size per tag, but runtime VRAM is weights plus KV cache and overhead, so you need headroom beyond the file (Ollama qwen2.5 tags).

Model (Q4_K_M default)	Size on disk	Fits 6 GB?	Fits 8 GB?	Fits 12 GB?
llama3.1:8b	~4.7 GB (tags)	Tight, unmeasured	Yes, measured	Yes
qwen2.5:7b	~4.7 GB (tags)	Tight, unmeasured	Yes, measured	Yes
qwen2.5:14b	~9 GB (tags)	No	No, spills	Should fit, unmeasured on a 3060

Expect 70+ tokens/s from 8 GB on 7B/8B models

We measured this, and 8 GB is plenty for this class. With a 150 W software cap, the RTX 3060 Ti ran qwen2.5:7b at 76.3 tokens/s generation and 3,847 tokens/s prompt processing, at 149.6 W and 51 C. llama3.1:8b ran at 73.9 tokens/s, 149.9 W, 51 C (bench notes).

Method: Ollama 0.34.2, Q4_K_M, one /api/generate call with stream:false and num_predict 320 after a warm-up. Tokens/s is eval_count/eval_duration, single runs, so treat ±5% as noise. Idle board power was 16.8 W (driver-reported).

Tip: The 3060 Ti is rated at 200 W (NVIDIA); we capped it at 150 W and still got 76 tokens/s. Cap first before you shop for a bigger PSU.

Pay for 12 GB if you want 14B, because spilling costs a third of your speed

If 14B models are the goal, 8 GB is a trap. Our qwen2.5:14b run split 6.7 GB on the 3060 Ti, 1.7 GB on the RTX 2060 and about 0.5 GB on CPU: 24.9 tokens/s. With the whole 2060 free, the same model sat fully in VRAM across both cards at about 38 tokens/s (bench notes).

Only the RTX 3060 12 GB tops 8 GB below the 3070 in NVIDIA’s 30-series table (NVIDIA compare). The RTX 4060 Ti also comes in 16 GB (NVIDIA), but our sources have no 16 GB price. The 3060 12 GB uses a 192-bit bus vs 256-bit on the 3060 Ti (NVIDIA), so we’d expect lower tokens/s than our 3060 Ti numbers; that’s our expectation, not a measurement.

Example (illustrative): You want qwen2.5:14b for coding help. At ~9 GB it spills on any 8 GB card, landing near our 24.9 tokens/s split result or worse. A 12 GB card should keep it fully on GPU.

Skip new 8 GB cards at $500+, and buy 12 GB refurbished

The old “$280 RTX 3060 12 GB” advice is dead, but refurbished units still undercut new 8 GB cards. On September 25, 2026, Newegg listed a refurbished MSI Ventus 2X OC 12G at $359.99 and a refurbished EVGA XC at $394.99, versus new 3060 12 GB cards from $459.97 (Newegg). Amazon Renewed 3060 12 GB listings ran $399.97–$429.97, mostly 1–3 in stock, and new ones $457.58–$539.99 (Amazon).

The same day, a new RTX 4060 8 GB was $530.00+ on Amazon (Amazon) and $499.99–$543.36 on Newegg (Newegg). A new RTX 5060 8 GB was $539.99 (Amazon). That’s $140–$180 more for less VRAM.

Watch out: New RTX 2060 6 GB listings ran $259–$369.99 (Amazon). At $339–$370 you’re a few dollars from a refurbished 12 GB 3060. Only consider a 2060 on used offers, which started at $185.99.

The Intel Arc B580 12 GB listed at $329.99 new (Amazon), but it’s outside this NVIDIA guide and we haven’t tested it.

Run this worksheet before you click buy

1. Write down the biggest model you’ll actually run and its default tag size (qwen2.5, llama3.1).
2. Add headroom for KV cache and overhead. Our rule: a 14B Q4 needs about 9 GB to stay on GPU.
3. If that number is at or under ~6 GB, an 8 GB card works. If it’s around 9 GB, you need 12 GB.
4. Price that VRAM tier refurbished first, then used, then new. Note stock counts; renewed listings were 1–3 deep.
5. Check your PSU: NVIDIA lists 550 W system power for the 3060 and 600 W for the 3060 Ti (NVIDIA).
6. Plan a power cap. Ours kept 70+ tokens/s at 150 W on the 3060 Ti.
7. Confirm the exact variant: the RTX 3060 also exists as an 8 GB, 128-bit card. The 12 GB model is the one you want.

What we couldn’t verify: we haven’t measured an RTX 3060 12 GB, RTX 4060 or a 6 GB card running a model alone, we have no RTX 3060 Ti price in our sources, and we can’t vouch for refurbished warranty terms or card condition from listings alone.