Dual RTX 3090 (used)

The community sweet spot nobody will put on a product page.

Two used 3090s give you 48 GB of VRAM for roughly $1,600 all-in — enough for 70B dense at Q4 with room for context. llama.cpp and Ollama split across PCIe automatically; no NVLink needed. The compromise is noise, heat, and finding honest used cards.

The decision in five lines

The call: Buy — The community sweet spot nobody will put on a product page.
Best for: Smart money
Runs well: Qwen3-Coder-30B-A3B (MoE, fits 24GB) · Qwen 3.6-27B · Z-Image-Turbo (Apache 2.0)
Watch out: Used 3090 floor lifted in May 2026 to $950–$1,200 on eBay (median ~$1,050) as buyers priced out of 5090 scarcity moved down a tier. The dual-3090 build now costs ~$2,100 for two cards plus PSU upgrade — still the cheapest path to 48 GB, but closer to dual-card 4090 territory than it was in April.
Evidence: Estimated · last verified June 2026

2×24: GB GDDR6X
936: GB/S BANDWIDTH
700: W COMBINED TDP
~$1,800: ALL-IN (USED)

What fits at this tier

Effectively matches the RTX 5090 for any model that fits in 48 GB — which is everything up to 70B Q4. MoE picks run slightly slower than single-card due to PCIe split, but the 48 GB headroom unlocks longer context windows than a 5090 can hold.

CODING

Qwen3-Coder-30B-A3B (MoE, fits 24GB) Community daily driver for local coding; 3B-active MoE delivers 30B quality at 3B-dense speed.

CHAT / GENERAL

Qwen 3.6-27B April 22 2026 dense refresh; supersedes Qwen 3.5 27B and claims to beat the prior 397B MoE flagship while staying single-GPU at Q4 (~17 GB).

DOCS & RETRIEVAL

Qwen 3.6-27B April 22 2026 dense refresh — 262K native context extensible to 1M, multimodal, single-GPU at Q4. Now the dense long-context top pick.

IMAGE

Z-Image-Turbo (Apache 2.0) Community daily driver for realism; 6B, 8-step inference, Apache 2.0 — commercial OK.

AGENTS

Qwen 3.6-35B-A3B Latest Qwen MoE; strong function calling; realistic on 24GB+ VRAM or Mac 48GB+ — the local agentic top pick.

VOICE

Qwen3-Omni-30B-A3B-Instruct Apache 2.0 MoE; audio+video+image+text in, speech+text out; 17GB at Q4. Frontier unified voice.

The call

Buy it if you value VRAM per dollar over every other metric, have a case that handles two 350 W cards, and don't mind spending an evening testing used GPUs on arrival.
Skip it if you want silence, a small form factor, or the option to sell the setup easily in 3 years. Also skip if you can't find 3090s under $1,000 locally — at $1,200+ each the math stops working.

Watchouts

Used 3090 floor lifted in May 2026 to $950–$1,200 on eBay (median ~$1,050) as buyers priced out of 5090 scarcity moved down a tier. The dual-3090 build now costs ~$2,100 for two cards plus PSU upgrade — still the cheapest path to 48 GB, but closer to dual-card 4090 territory than it was in April.
Combined 700 W TDP needs a 1,200 W+ PSU with two 8-pin EPS cables. Plan on $200–$400 PSU upgrade if coming from a single-card build.
Two cards need real airflow. A mid-tower case with bottom intake + top exhaust works; slim cases or vertical mounts cause thermal throttling fast.
Loud. Both cards hit 75–80°C fan curves under sustained load. Plan for fan noise like a dual-GPU gaming rig, because it is one.

Local vs cloud at this tier

● LOCAL WINS

Best $/GB-VRAM in the entire market. Handles every 70B Q4 and MoE 35B-A3B model the RTX 5090 does, often for less than half the price.

● CLOUD WINS

Cloud wins on power bills at light usage, and on first-day model access. If you're running <10 hours/week, a $20/mo Claude Pro plan is cheaper than the electricity.

At roughly $1,600 all-in with regular usage, break-even vs a $20/mo Claude Pro plan is ~24–36 months depending on your power cost. Heavy coding/agent users break even in 6–12 months.

Next step

Load this setup into the planner→