A budget GPU for local LLMs means 12 GB or 16 GB of VRAM. At Q4_K_M, a 14B model needs about 10.4 GB (14 × 0.5 × 1.2 + 2), so a 12 GB card runs it. The RTX 5070 is a 12 GB card, not 16 GB. A 16 GB card such as the RTX 5070 Ti fits a 20B at about 14.0 GB, at a short context. A 27B at 18.2 GB fits neither tier and needs a 24 GB card. Estimates as of 19 October 2026.
Editorial: AI handled the first pass on this budget round-up. The fit math and the final picks were checked against the cited sources before publishing.
Key facts
RTX 5070 is 12 GB
It is not a 16 GB card. Check current price on Amazon.
14B Q4_K_M
10.4 GB (14 × 0.5 × 1.2 + 2). Fits 12 GB. Qwen3.8-27B Q4 is 18.2 GB and does not.
Granite 4.2-8B
IBM Q4_K_M file is 5.35 GB. Formula is 6.8 GB. Fits 8 GB. The 30B Q4_K_M file is 17.72 GB and does not fit 16 GB.
Checked
2026-10-13
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Manufacturer MSRP (launch list price)
These are NVIDIA’s launch MSRPs from NVIDIA’s own pages, checked 17 October 2026. This is not Amazon’s price. Current prices are on Amazon, and partner cards often list higher.
Manufacturer MSRP (launch list price), checked 17 October 2026
Checked 17 October 2026. Formula matches the VRAM calculator.
Citeable facts
Key fact: The RTX 5070 is a 12 GB card. NVIDIA’s launch MSRP is in the manufacturer MSRP table on this page. This page does not show an Amazon price. Granite Switch 4.1 8B at Q4 is 7.3 GB and fits 8 GB and 12 GB. Dory 1.0 4B at Q4 is 4.4 GB and fits 8 GB. Granite Switch 4.1 30B at Q4 is 20.0 GB and does not fit 16 GB.
How we got it: NVIDIA GeForce news, February 2025, re-read 17 October 2026. 8.78 × 0.6 + 2 = 7.3. 3.98 × 0.6 + 2 = 4.4. 29.99 × 0.6 + 2 = 20.0. Hugging Face safetensors totals the same day.
Checked:
Key fact: A 14B model at Q4_K_M needs 10.4 GB and fits 12 GB. The same model at Q8 needs 18.8 GB and does not fit 16 GB.
Methodology
— parameter math, quantization bytes, and the source list.
Every budget tier has a clear winner — and a few traps to avoid. This guide covers the best GPU for local LLM inference at every budget, with benchmark speeds on Qwen3 14B Q4 and the exact reasons to pick or skip each card.
Checked listings, not a generic search. Check current price on Amazon fits qwen3:14b. Check current price on Amazon fits 20B Q4 (14.0 GB) and still misses 14B Q8. Check current price on Amazon for the RTX 5070 Ti 16 GB (B0DV9GMDLR). The PNY RTX 5060 Ti 16 GB (B0F68R489D, 448 GB/s) said currently unavailable on 17 October 2026. Granite 4.2-8B fits these 8 GB and 12 GB cards: IBM’s Q4_K_M file is 5.35 GB and the formula is 6.8 GB. Granite Switch 4.1 8B at Q4 is 7.3 GB and fits the same cards. Dory 1.0 4B at Q4 is 4.4 GB and fits 8 GB. Qwen3.8-27B Q4 is 18.2 GB and does not fit a 16 GB card. Granite Switch 4.1 30B at Q4 is 20.0 GB and does not either. Run ollama run granite4.2:8b on the budget tier and size it in the VRAM calculator. The sibling pages are the Qwen3.6 and Granite 4.2 fit table and the local coding guide.
Checked 17 October 2026. The 4060 page showed Buy New with only 1 left. The 4060 Ti page said currently unavailable. Check current price on Amazon.
Links to Amazon are affiliate links. No extra cost to you.
GPU
VRAM
Bandwidth
14B Q4 Speed
Price
New/Used
RTX 3060 12 GB
12 GB
360 GB/s
14B Q4 fits
Check price on Amazon
Used
Intel Arc B580 12 GB
12 GB
456 GB/s
14B Q4 fits
Check price on Amazon
New
RTX 5060 8 GB
8 GB
448 GB/s
N/A (8 GB)
Check price on Amazon
New
AMD RX 9060 XT 16 GB
16 GB
320 GB/s
20B Q4 fits
Check price on Amazon
New
RTX 5060 Ti 16 GB
16 GB
448 GB/s
14B Q4 fits
No featured offer (B0F68R489D)
New
RTX 5070 12 GB
12 GB
672 GB/s
14B Q4 fits
Check current price on Amazon
New
AMD RX 9070 XT 16 GB
16 GB
640 GB/s
n/a
Check price on Amazon
New
RTX 5070 Ti 16 GB
16 GB
896 GB/s
14B Q4 fits
Check current price on Amazon
New
RTX 4090 24 GB
24 GB
1,008 GB/s
127.7 tok/s Llama 3 8B (XiongjieDai)
Check price on Amazon
New
Speed figures are for Qwen3 14B Q4 on Linux with Ollama. Use the VRAM Calculator for exact memory requirements.
Budget Tier Breakdown
Entry tier
Used Market Pick
Best pick: RTX 3060 12 GB
Why: 12 GB at the entry tier. 360 GB/s bandwidth. Fits Qwen3 14B at Q4_K_M (10.4 GB).
Alternative: RTX 3080 10 GB — faster but less VRAM, worse for 14B models
Avoid: RTX 4060 8 GB — less VRAM than RTX 3060 12 GB for more money
RTX 3060 12 GB (used)
Check price on Amazon
VRAM
12 GB
14B Q4
14B Q4 fits
Price
Check price on Amazon
Pros
+ 12 GB on the used market
+ 12 GB fits Qwen3 14B at Q4
+ Full CUDA support
Cons
- Used market — no warranty
- 360 GB/s bandwidth is modest
- Older Ampere architecture
At the entry tier, the used RTX 3060 12 GB is the pick because it has 12 GB. On eBay or local resale markets, you get full CUDA support, and Qwen3 14B at Q4_K_M (10.4 GB) with about 1.6 GB to spare. The only caveat is the used market: inspect listings carefully and factor in no warranty.
Alternative: Used RTX 3060 12 GB if you prioritize CUDA ease over new hardware
Avoid: RTX 5060 8 GB — same price as Arc B580 but only 8 GB VRAM
Intel Arc B580 12 GB
Check price on Amazon
VRAM
12 GB
14B Q4
14B Q4 fits
Price
Check price on Amazon
Pros
+ 12 GB GDDR6, new with warranty
+ 456 GB/s vs 360 GB/s on the RTX 3060
+ Runs in Ollama on Linux
Cons
- Less mature CUDA ecosystem than NVIDIA
- Windows driver support still improving
- Some frameworks need manual setup
The Intel Arc B580 gives you 12 GB new, where the RTX 5060 has 8 GB. 12 GB fits Qwen3 14B at Q4_K_M (10.4 GB); 8 GB does not. We have not measured its tok/s. The trade-off is Intel's GPU software ecosystem, which is solid on Linux but lags NVIDIA on Windows for some tools.
Mid-budget tier
AMD 16 GB Pick
Best pick: AMD RX 9060 XT 16 GB
Why: 16 GB GDDR6 at 320 GB/s (AMD spec). Fits 14B (10.4 GB) and 20B (14.0 GB) at Q4_K_M.
Alternative: RTX 5060 Ti 16 GB if you need Windows CUDA ease
Avoid: RTX 4060 Ti 8 GB — never buy 8 GB when 16 GB exists at this price
AMD RX 9060 XT 16 GB
Check price on Amazon
VRAM
16 GB
14B Q4
20B Q4 fits
Price
Check price on Amazon
Pros
+ 16 GB VRAM
+ 160 W board power (AMD)
+ RDNA 4 architecture
Cons
- ROCm setup required for some tools
- CUDA ecosystem gaps vs NVIDIA
- Windows LLM app support varies
16 GB fits qwen3:14b at Q4_K_M (10.4 GB) and a 20B model at Q4_K_M (14.0 GB). It does not fit 14B at Q8 (18.8 GB) or 32B at Q4 (21.2 GB). The PNY RTX 5060 Ti 16 GB (B0F68R489D) is 448 GB/s and had no featured Amazon offer on 2 October 2026. The command is ollama run qwen3:14b. In LM Studio, use the Q4_K_M GGUF, not Q8.
Value tier
16 GB CUDA Pick
Best pick: Used RTX 4060 Ti 16 GB or AMD RX 9060 XT 16 GB
Why: 16 GB VRAM, 288 GB/s bandwidth for RTX 4060 Ti — lower bandwidth than RX 9060 XT but CUDA works everywhere.
Alternative: RX 9060 XT 16 GB has better bandwidth if CUDA is not a requirement
Avoid: RTX 5060 Ti 8 GB — avoid any 8 GB card when 16 GB is available near this price
RTX 4060 Ti 16 GB (used)
Check price on Amazon
VRAM
16 GB
14B Q4
20B Q4 fits
Price
Check price on Amazon
Pros
+ 16 GB CUDA VRAM at a low used-market price
+ Works with every NVIDIA-optimized tool
+ Low 165 W TDP
Cons
- 288 GB/s bandwidth, the lowest here
- Speed not measured here
- Used market — check condition
RTX 5060 Ti 16 GB
Check price on Amazon
VRAM
16 GB
14B Q4
20B Q4 fits
Price
Check price on Amazon
Pros
+ 16 GB GDDR7, Blackwell architecture
+ Full CUDA support out of the box
+ New with warranty
Cons
- 448 GB/s on the PNY 16 GB listing (B0F68R489D), not ~480
- No featured Amazon offer on 2 October 2026 — no current price
At this tier you have two strong paths to 16 GB VRAM. The used RTX 4060 Ti 16 GB is the CUDA play — every tool works, no setup friction, and the price is good. The AMD RX 9060 XT 16 GB new matches VRAM and has a little more bandwidth (320 vs 288 GB/s, AMD and NVIDIA specs). We have not measured the speed gap. If you need fine-tuning or Windows app compatibility, the RTX 4060 Ti or the new RTX 5060 Ti 16 GB is the safer pick.
Mainstream tier
Blackwell Value
Best pick: RTX 5070 12 GB
Why: 12 GB GDDR7 at 672 GB/s. Fits qwen3:14b at Q4_K_M (10.4 GB), not 32B. Check current price on Amazon.
Alternative: Check current price on Amazon
Avoid: Calling this card 16 GB — the catalog id is nvidia-rtx-5070-12gb
RTX 5070 12 GB
Check current price on Amazon
VRAM
12 GB
14B Q4
14B Q4 fits
Price
Check current price on Amazon
Pros
+ 12 GB GDDR7, 672 GB/s — not a 16 GB card
+ Fits ollama run qwen3:14b (10.4 GB at Q4_K_M)
+ Check current price on Amazon
Cons
- 12 GB does not fit qwen3:32b (21.2 GB) or 14B Q8 (18.8 GB)
- Check current price on Amazon
- A 16 GB card is the step up if you want 20B Q4 (14.0 GB) with room
The RTX 5070 is a 12 GB card. 12 GB fits qwen3:14b at Q4_K_M (10.4 GB). It does not fit 14B at Q8 (18.8 GB) or qwen3:32b (21.2 GB). In LM Studio, download the Q4_K_M GGUF of Qwen3 14B. Run ollama ps after load; GPU% under 100 means the model spilled to system RAM. Check current price on Amazon.
Upper tier
16 GB Blackwell Pick
Best pick: RTX 5070 Ti 16 GB
Why: 896 GB/s bandwidth (NVIDIA), Blackwell, full CUDA. Fits 20B at Q4_K_M (14.0 GB); does not fit 32B at Q4 (21.2 GB).
Alternative: AMD RX 9070 XT 16 GB — 16 GB at up to 640 GB/s (AMD spec) if CUDA is not needed
Avoid: RTX 4090 — overkill if you do not need 24 GB VRAM
RTX 5070 Ti 16 GB
Check price on Amazon
VRAM
16 GB
14B Q4
20B Q4 fits
Price
Check price on Amazon
Pros
+ 896 GB/s GDDR7 (NVIDIA)
+ Fits 14B and 20B at Q4_K_M
+ Blackwell architecture, full CUDA, new with warranty
Cons
- Still 16 GB — cannot run 32B at full Q4
- Premium over RTX 5070 for more bandwidth
- 300 W TGP; NVIDIA asks for a 750 W PSU
The RTX 5070 Ti 16 GB is the recommendation for most LLM users who want fast, high-quality local AI without paying flagship prices. Its 896 GB/s GDDR7 bandwidth is the same as the RTX 4090 on models that fit in 16 GB — and the price gap versus the RTX 4090 is hard to justify unless you specifically need 24 GB for 32B models. For everyday 14B inference, this card delivers the best experience per dollar in 2026.
What VRAM Tier Do You Actually Need?
For budget LLM use, 12GB VRAM is the sweet spot in 2026 — it runs 7B models at Q8 and 13B at Q4_K_M comfortably. The RTX 4070 12GB and Intel Arc B580 12GB are the top value picks. Avoid 8GB cards if you plan to run anything larger than 7B.
8 GB
7-8B models only. Fine for casual use, limiting long-term. Avoid in 2026 if you can stretch to more VRAM.
12 GB
14B Q4 fits. Good all-around. Best value tier — the Arc B580 and RTX 3060 hit this sweet spot.
16 GB
14B Q4 (10.4 GB) and 20B Q4 (14.0 GB) fit. 14B Q8 is 18.8 GB and does not. 27B Q4 is 18.2 GB and does not.
24 GB
32B Q4 fits. Power user territory. Used RTX 3090 is the budget path; RTX 4090 for speed.
48 GB+
70B Q4 fits. Serious research use. Mac Studio M4 Max (64 GB unified) or dual RTX 4090s.
The "Never Buy" List for LLMs
These GPUs are not bad for gaming — but for local LLM inference, they represent poor value. Better options exist at the same or lower prices.
RTX 5060 8 GB
Check price on Amazon
Same price as Arc B580 but half the VRAM — 8 GB in 2026 is a dead end
RTX 4060 8 GB
Check price on Amazon
8 GB at 272 GB/s — does not fit 14B at Q4 (10.4 GB)
RTX 4060 Ti 8 GB
Check price on Amazon
Avoid — the 16 GB version is available at similar or only slightly higher used prices
Frequently Asked Questions
What is the best budget GPU for LLMs?
The Intel Arc B580 is our new-card budget pick. It has 12 GB GDDR6 and 456 GB/s bandwidth, which fits Qwen3 14B at Q4_K_M (10.4 GB). We have not measured its tok/s. The RTX 5060 has 8 GB, which does not fit 14B at Q4. On Linux, Ollama supports Intel Arc natively.
What is the best budget GPU for 16 GB VRAM?
A 16 GB card fits qwen3:14b at Q4_K_M (10.4 GB) and a 20B model at Q4_K_M (14.0 GB). It does not fit 14B at Q8 (18.8 GB) or 27B at Q4_K_M (18.2 GB). The PNY RTX 5060 Ti 16 GB (B0F68R489D) is 448 GB/s. On 2 October 2026 Amazon had no featured offer, so there is no current price. The command is ollama run qwen3:14b.
Is 8 GB VRAM enough for local LLMs in 2026?
8 GB fits qwen3:8b at Q4_K_M (8 × 0.5 × 1.2 + 2 = 6.8 GB). A 14B model at Q4_K_M is 10.4 GB and does not fit. Q8 of an 8B model is 11.6 GB and needs 12 GB. The command is ollama run qwen3:8b.
Should I buy a used or new GPU for LLMs?
At the entry tier, used (RTX 3060 12 GB) is the best option — nothing new at that budget matches the VRAM. At higher budget tiers, new GPUs (Arc B580, RX 9060 XT, RTX 5070 Ti) offer modern architecture, warranty, and better efficiency. A used RTX 3090 is the low-cost path to 24 GB (32B at Q4 is 21.2 GB).
VRAM figures on this page come from the site’s VRAM formula and model cards. The sitewide formula and the full source list are on the methodology page. For this guide we used:
Spot a number that does not match the linked source? Contact us and we will update the guide.
We use cookies and personalised ads
This site uses cookies and shows personalised ads via Google AdSense. We and our partners store and access information on your device to serve relevant ads and improve your experience.
You can accept all cookies, decline (non-personalised ads only), or
manage preferences.
See our Privacy Policy.
Cookie preferences
Choose which cookies you allow. Strictly necessary cookies are always active.
Strictly necessary
Session state, security, and performance. Cannot be disabled.
RTX 3060 12 GB
RTX 5060 8 GB
AMD RX 9060 XT 16 GB
RTX 4090 24 GB