What is the cheapest hardware that runs a 70B model locally at usable speed?

As of August 2026, two used RTX 3090s at $2,000-2,600 total. The 48GB of combined VRAM holds a dense 70B at Q4_K_M with context headroom, and community benchmarks put generation at 16-22 tokens per second. Every cheaper option either fails to fit the model or generates below reading pace.

How much memory does a 70B model need?

Roughly 40GB for the weights at Q4_K_M quantization, plus KV cache that grows with context length. Practically you want 48GB of VRAM, or 64GB+ of unified memory, for comfortable daily use. At Q8 the weights roughly double to about 75GB, which pushes you to 96GB-class machines.

Can a 128GB Ryzen AI Max 395 mini-PC run a 70B?

It fits easily, but a dense 70B at Q4 generates around 5 tokens per second, because the roughly 215 GB/s of measured memory bandwidth is the ceiling. Token generation is bandwidth-bound: the chip must read all active weights for every token. These boxes are excellent for MoE models with small active parameter counts, and poor for dense 70B models.

Do I still need a dense 70B in 2026?

Ask before you spend. The strongest open models of 2025-2026 moved to MoE architectures (gpt-oss 120B has about 5B active parameters) and to strong 27-35B dense models. Many workloads that needed a 70B in 2024 are now better served by either — and both run on much cheaper hardware. The 70B tier is worth buying for specifically dense-70B fine-tunes and evaluation work.

Is a Mac the cheapest way to run a 70B?

No longer. Apple removed the 256GB and 512GB Mac Studio configs in 2026, and the surviving M3 Ultra starts at $5,299 with a 96GB ceiling. That runs a 70B at Q8 well, but at more than twice the price of a dual RTX 3090 build. The Mac case is silence, warranty, and low power — not price.

← All guides

Hardware August 11, 2026

The Cheapest Way to Run a 70B Model Locally in 2026

Our site has five separate pages about whether some machine can run a 70B model. This page answers the question those pages set up: what is the cheapest box that actually does it? The trap in 2026 is that 'fits' and 'runs' have diverged. A $2,000 mini-PC with 128GB of unified memory holds a 70B easily — and generates about 5 tokens per second. Ranked below by price, with the honest speed attached to each.

Bottom Line

Cheapest that actually runs one: two used RTX 3090s, $2,000-2,600 (as of August 2026), 16-22 tok/s on 70B Q4. Loud, ~700W, no warranty — but nothing cheaper clears reading pace.
Cheapest that merely fits one: a 128GB Ryzen AI Max+ 395 box (~$2,000-2,400). Dense 70B at Q4: ~5 tok/s. Fits ≠ runs.
The quiet-and-warrantied route: Mac Studio M3 Ultra 96GB, from $5,299. The 256GB/512GB configs are discontinued — plan around 96GB.
Before spending anything: check whether you still need a dense 70B at all. The 2026 model landscape moved to MoE and strong 27-35B models, and both run on far cheaper boxes.

The number that decides everything

A dense 70B reads ~40GB of weights for every token it generates. Speed is therefore memory bandwidth divided by model size, and no tuning escapes it. That is why the ranking below sorts strangely: capacity is cheap in 2026, bandwidth is not.

Route	Price (Aug 2026)	Memory	Bandwidth	Dense 70B Q4 speed	Verdict
2x used RTX 3090	$2,000-2,600	48GB VRAM	~936 GB/s/card	16-22 tok/s	The pick
Ryzen AI Max+ 395 128GB box	~$2,000-2,400	128GB unified	~215 GB/s measured	~5 tok/s	Fits, crawls
Used RTX A6000 48GB	$2,600-3,800	48GB VRAM	~768 GB/s	~15-20 tok/s (est.)	Single-slot sanity, worse $/GB/s
NVIDIA DGX Spark	$4,699	128GB unified	273 GB/s	bandwidth-bound, ~5-8 tok/s (est.)	Buy it for CUDA, not for 70B speed
Mac Studio M3 Ultra 96GB	from $5,299	96GB unified	~800 GB/s	70B Q8 at usable speed	Best quiet option, not cheapest
RTX PRO 6000 Blackwell 96GB	$13,250	96GB VRAM	very high	fast, single card	Wrong audience for this page

Estimates marked (est.) are derived from the bandwidth-divided-by-weights ceiling, not from a measured run — treat them as caps, not promises. All prices are street ranges as of August 2026; the memory shortage moves them monthly.

Route 1 — dual used RTX 3090s: still the answer, at a worse price

The classic recipe survived the 2026 price spike, barely. Used 3090s now trade at $1,000-1,300 each (not the $650-750 that older guides quote), so the pair costs $2,000-2,600. That is still half of everything else that clears 15 tok/s. The full tradeoff — power, the no-NVLink truth, the 5090 counter-case — is in dual RTX 3090 vs RTX 5090, and each card should pass the used-3090 checklist before its return window closes.

The honest costs: ~700W of GPU under load, real noise, two chances at a dud card, and zero warranty. At ~$0.18/kWh (2026 US average) heavy use adds tens of dollars monthly — break-even math here.

Route 2 — the 128GB unified-memory trap

The Ryzen AI Max+ 395 boxes are the best-value machines we cover for MoE models — our model picks for them show gpt-oss 120B at 31-55 tok/s, because only ~5B parameters are active per token. But point one at a dense 70B and the ~215 GB/s measured bandwidth caps you near 5 tok/s. Same story for the DGX Spark at 273 GB/s, at twice the price ($4,699 after NVIDIA’s official $700 “memory supply constraints” increase — our Spark verdict covers whether it earns its keep another way).

If a seller pitches you a 128GB box “because it runs 70B models,” ask them at what speed. Capacity marketing leans on exactly this confusion.

Route 3 — used RTX A6000: paying for one slot

A used A6000 48GB ($2,600-3,800, wide condition spread — enterprise lease returns) does what the dual 3090s do in one slot, one blower, ~300W. You pay roughly $600-1,200 extra for simplicity and get slightly less bandwidth. Reasonable if your case or PSU cannot take two cards; otherwise the pair wins. The 48GB VRAM + 128GB RAM page covers what this tier does with big system RAM behind it.

Route 4 — the Mac, now that the big configs are gone

The old advice — “buy a 256GB Mac Studio” — died in 2026: Apple withdrew the 512GB (March) and 256GB (May) M3 Ultra configs. The surviving ceiling is 96GB, from $5,299. With ~800 GB/s of bandwidth that runs a 70B at Q8 at genuinely usable speed, silently, under warranty, at a fraction of the power. It is the best machine on this page to live with, and at 2-2.5x the dual-3090 price it is not the cheapest way to do anything.

The question to ask before any of it

Do you still need a dense 70B? The open-model frontier moved: MoE models (gpt-oss 120B, ~5B active) deliver frontier-adjacent quality at small-model bandwidth cost, and 27-35B dense models absorbed most agentic work. If your workload is agents or coding, a single 24GB card and the right 27B-class model likely serves you better than any box on this page. The 70B tier earns its cost for dense-70B fine-tunes, evaluation work, and specific model requirements — know which buyer you are before spending $2,000+.

The hardware from this page:

EVGA RTX 3090 24GB — two of these is the cheapest real 70B rig.
AMD Ryzen AI Max+ 395 128GB box — the MoE-era alternative, if you don't need dense 70B.
NVIDIA DGX Spark 128GB — the CUDA dev box, bought for the stack rather than the speed.
NVIDIA RTX PRO 6000 Blackwell 96GB — the money-no-object single card.

Amazon affiliate links — we earn a small commission at no cost to you.