What $1,000 buys for local LLMs (Sep 2026)

As of 2026-10-07. Prices are our 2026-09-28 snapshot; every row links to its source. The B60 row is unchanged since 2026-09-10 because we have no newer reading for it. This is a VRAM-first guide, because VRAM is what decides which models you can run at a usable speed.

The three cards in this budget

That is the whole new-card ladder under $1,000 as of the 2026-09-28 snapshot: 12GB ($495), 16GB ($800), 24GB ($959) — and 32GB ($949.99). The 32GB Intel Arc Pro B70 changed this guide's conclusion: until it appeared, the cheapest way past 24GB was a used card or a much larger budget.

What each tier actually runs

A rule of thumb for local inference at 4-bit: weights cost roughly 0.6 GB per billion parameters, plus 0.1–0.2 GB per billion for KV cache and runtime overhead (both grow with context length). That gives:

• 12GB (RTX 3060) — 7–8B models at Q4 comfortably; 13–14B at Q4 with a short context. This is the cheapest way into the category, and it is why the old 3060 keeps showing up in entry builds.
• 16GB (RTX 5060 Ti) — 13–14B at Q4 with room for context, or 8B at Q8. A more comfortable daily driver than 12GB.
• 24GB (Arc Pro B60) — the 27–32B class at Q4, or a 30B-class MoE with the experts held in VRAM. This is the tier where "a genuinely useful local model" starts.
• 32GB (Arc Pro B70) — the same class with real context headroom: a 27B at Q4 plus a long KV cache, or a 70B-class MoE with heavy offload. Our board has 16 sourced runs on this card, best 283.1 tok/s.
• 48GB and up — 70B-dense at Q4. That is a different budget (A6000 48GB is $4,849.99 refurbished right now), and why dual-used-3090 builds stay popular.

The used-market wildcard

A used RTX 3090 (24GB) is the classic mid-budget move, but its price has moved with the shortage: $1,300 in the 2026-09-28 snapshot, up from ~$1,050 in May. At $1,300 it buys 24GB for $54/GB, still the best VRAM-per-dollar we track — the live table is here.

Look at throughput per dollar, not FLOPS

Every card above has benchmark rows on our board, filtered to sourced runs: RTX 3060 · RTX 5060 Ti 16GB · Arc Pro B60 · RTX 3090. Compare like-for-like pairs on the RTX 3090 vs RTX 5090 page, and see VRAM-per-dollar plus tokens-per-dollar on the /value index. Remember that "best tok/s" on a card is measured on some specific model — check the engine, quant and context on each row before you generalise.

FAQ

How much VRAM do I need for a 27B model at Q4?

Around 16–18GB for the weights at 4-bit, plus KV cache — so realistically a 24GB card if you want a usable context length. A 16GB card can load a 27B Q4 but leaves very little room for context.

Is the RTX 3060 12GB still worth buying in 2026?

As the cheapest new 12GB card at $495 (2026-09-28), yes — for 7–8B models at Q4 it is the lowest-cost entry point we track. Its limit is context: 12GB does not leave much room once the weights are loaded.

Do I need 48GB for a 70B model?

For 70B-dense at Q4, yes — roughly 40GB of weights plus cache. 24GB cards can run 70B only with heavy offload, which costs speed. If 70B is the goal, plan for the 48GB tier.

Is a used RTX 3090 better than a new 16GB card?

For VRAM per dollar, usually — 24GB used versus 16GB new, if you can source a clean card. The trade-offs are warranty, power draw and noise. Check both against the value index before deciding.

Sources

• RTX 3060 12GB: Newegg — MSI Ventus RTX 3060 12GB
• RTX 5060 Ti 16GB: Newegg — MSI SHADOW RTX 5060 Ti 16GB
• Arc Pro B60 24GB: Newegg — GUNNIR Intel Arc Pro B60 BS 24GB
• RTX A6000 48GB: Newegg — refurbished
• New-card prices ($495 / $800 / $959 / $949.99) and used RTX 3090 ($1,300): 2026-09-28 snapshot
• Used RTX 3090 market history (May 2026, ~$1,050): Howard Resource Group
• Live data: /prices · /value · /benchmarks · methodology