DiggTop
Menu
⌕

Local LLM GPU Buying Guide (September 2026)

As of 2026-09-28 for the cards in our weekly snapshot (US, new unless marked "used"); cards we have not re-snapshotted since carry their original date in the source list. And they move daily. 2026 is a seller's market: the memory shortage that pushed global PC shipments down 4.9% in Q2 (IDC) is inflating GPU prices month over month — TechSpot measured roughly +15% across the market in August alone. Re-check prices before you buy, and treat this guide's tiers as the durable part, not the dollar figures.

TL;DR — picks by VRAM tier

What you want to run VRAM you need Our pick (Sep 2026) Why
7B–9B (chat, coding assist) 8–12 GB RTX 3060 12GB (new, ~$495, 2026-09-28 snapshot) NVIDIA relaunched it in 2026 at $329; demand pushed it up ~50% in two months. Still the community workhorse (hundreds of LocalMaxxing runs on 12GB cards).
13B–32B (Q4) 16–24 GB Used RTX 3090 24GB (~$1,300, 2026-09-28 snapshot) eBay-sold baseline was ~$1,050 in May 2026; the used tier has climbed steadily since. 24GB is the single most submitted rig class in the community.
24GB brand-new 24 GB Intel Arc Pro B60 24GB ($599) The cheapest new 24GB card you can buy; workstation-grade drivers, no resale silliness.
70B class (Q4, ~40GB) 48 GB 2× used RTX 3090 (~$2,000–2,600) or used RTX A6000 48GB (~$2,200–2,800) Dual 3090 is cheaper and faster for batch work; A6000 wins on simplicity, power and NVLink.
32GB brand-new flagship 32 GB RTX 5090 (~$9,000 US street average, 2026-09-28 snapshot) — overpriced for most; AI PRO R9700 32GB ($1,329) if you need new workstation Yes, the 5090 is a 32GB card at a 4.5× street premium. For 70B Q4 you still need 48GB — see the tier above.

Why VRAM is the whole game

Local inference speed is dominated by one number: does the model fit in VRAM, and with how much context?

A rough planning rule for Q4 quantization:

This is why 24GB cards — new or used — are the price anchor of the whole 2026 market, and why the used market for 48GB workstation cards refuses to die.

What 2026's market looks like right now

Sourced market facts (sources at the bottom):

The takeaway: this is a market for decision frameworks and weekly price checks, not for "buy once, done" reviews.

Tier by tier, in detail

Tier 1: entry (8–12GB) — 7B–9B models

Tier 2: sweet spot (24GB) — 13B–32B Q4, up to 27–32B with low quant

Tier 3: 70B Q4 (40–48GB) — the "real models" tier

Tier 4: brand-new 32GB and beyond

Used-market watch rules

What to do next

  1. Decide the biggest model you will run in the next 12 months — that fixes your VRAM tier.
  2. Watch prices for a week (we publish weekly snapshots) — in this market a week can move a used 3090 by $100.
  3. Buy used-but-proven from sellers with history; ask for memory tests.

DiggTop is reader-supported: some links in future editions may be affiliate links, always disclosed. Benchmarks cited from community sites carry their own methodology — verify before big purchases.

Verify before you buy — DiggTop data tools

Don't take our word for it. Check the live data first:

Related reading: The VRAM Cliff · Run a 70B on two used RTX 3090s · $800 entry build · $1,500 sweet-spot build.

FAQ

How much VRAM do I need for a 70B model?

A 70B at Q4_K_M is roughly 40–48 GB of weights plus KV-cache overhead — that is 48 GB territory: a used A6000, a pair of 24 GB cards, or the RTX PRO 6000. On 24 GB you are realistically running 24–32 B models at Q4. For 70B, buy VRAM, not speed — see the 70B on two used 3090s guide.

Is a used RTX 3090 still the best value in 2026?

For most buyers, yes on throughput: 24 GB at ~$1,300 used (2026-09-28 snapshot) still outruns new cards at the same model size, and no new card near that price holds 24 GB. If $/GB is what you are optimising, Intel's Arc Pro B60 at $599 is the cheaper route to 24 GB. The risks are power draw, cooler wear, and single-fan thermal history — check the used-market rules before paying.

Should I wait for next-generation cards?

Only if you are waiting for a memory configuration you specifically need. Generational speed bumps rarely change which models fit — the VRAM cliff does. If you need local inference now, buy for the model size you actually want to run.

Does NVFP4 on 50-series cards change the math?

For 50-series cards, yes — it is the main reason to pay the premium: NVFP4 runs some models at near-FP8 quality in roughly half the footprint. It is engine- and model-specific, though, so check the compatibility matrix before assuming it applies to your stack.

Sources

Last reviewed: 2026-09-04. Prices as of that date; they are not guaranteed stable.

↑