What GPU Do You Need to Run the Big Three at Home? VRAM Tiers, Checked November 2026

Home Rig Updated 2026-11-04 · 3 min read · #gpu #vram #rtx #hardware

Status board (checked 2026-11-04)


RTX PRO 6000 (96GB)	listed ~$9,990 (PCMag, Oct 12); street ~$11,000 with a 12+ week lead (UnifiedTrend, Nov 2)
RTX 5090 (32GB)	MSRP $1,999 → reported street ~$3,300 (TechPowerUp via UnifiedTrend, Nov 2) — supply capped by GDDR7
RTX 5080 (16GB)	MSRP $999 → reported street ~$1,950 (same source)
Why prices climb	TrendForce (Sep 30): GDDR7 output is limited because fabs tilt advanced process toward server DRAM; GDDR6 is shrinking out of the market
4Q26 memory outlook	general DRAM contract +10–15% QoQ, NAND +15–20% (TrendForce) — the card is half memory purchase now

What it is

For MoE models like the Big Three, the GPU’s job is narrow: hold the active slice and feed it fast. Total size lives in RAM and SSD. That means two numbers decide your card:

VRAM capacity — how much of the pack can sit on-card instead of in RAM;
Bandwidth — active params × bandwidth ≈ decode speed, which is why a used 3090 still beats a new midrange card.

Where each model actually runs (measured, not theoretical)

Model	Pack (weights in RAM+VRAM)	Card that was measured	Speed
Qwen3.8-Flash-Next (125B, 6B active)	Q2_0 ≈ 37.6 GB	1× RTX 5070 12GB + 64 GB DDR5	94 tok/s short chat, 76 at 128K ctx
GLM-5.3-Flash (320B, 18B active)	Maya-XS ≈ 73.4 GB	1× RTX 5070 12GB + 4-channel DDR4	16.6 tok/s
〃	〃	2× 5090 + 1× 4090 (88GB VRAM)	148.6 tok/s at 32K
MiniMax M3.1 (428B, 23B active)	Maya-S ≈ 122 GB	1× RTX PRO 6000 96GB + 253 GB RAM	48.6 tok/s at 256K (with DSpark drafter)
DeepSeek V4.1-Flash (748B)	Maya-XS 277 GB w/ Engram	1× RTX 4090 24GB + 192 GB DDR5	44.1 tok/s at 64K (Maya user report, issue #503)

Getting started (3 steps)

Pick the model, then read its pack size — not its parameter count. The RAM page has the same table from the other side.
Match the VRAM tier: 12GB = the Flash-Next class at 2–3 bit + fast RAM; 24GB = comfortable GLM-XS; 32GB = a whole Coder build on-card; 96GB = MiniMax alone; below that, add cards or lean on RAM.
Check the lead time before the price: the PRO 6000 class runs 12+ weeks out. If the card you want is back-ordered, a 2-card budget setup is the same money.

Best use

Card class	Right call when…
Used 12GB (3060/5070 class)	you want the trend’s face (Flash-Next) and accept 50–90 tok/s with good RAM
24GB (4090/3090)	GLM-XS at 128K ctx is the target; still the used-market value king
2× 24–32GB	the 148 tok/s GLM rig — agentic work gets actually usable
RTX PRO 6000 96GB	one-card MiniMax M3.1 at 48 tok/s — the measured box still had 253 GB of RAM behind it, and the waitlist runs 12+ weeks

Honest weaknesses

2026 prices move against you weekly; the board above is dated for a reason.
A big card without matching RAM is wasted money — RAM does the storage, the card does the math.
Speeds are hardware-specific: same card, different SSD or half the channels, different number. Bench your own box.

FAQ

Is a 5090 worth it over a 4090?

32GB vs 24GB and much more bandwidth — but at Nov-2026 street prices (~$3,300 vs a used 4090) you’re paying roughly the price of a 4090 extra for the memory. For MoE rigs, 2× 4090 usually beats 1× 5090.

Do I need a datacenter card?

Not for the Flash class — that’s the entire point of the RAM-offload trick. Datacenter cards only become necessary above ~400B if you dislike slow first tokens.

What about AMD or Intel?

Strata covers AMD and Arc; Project Maya is NVIDIA-first with experimental AMD. The measured numbers on this page are all NVIDIA.

📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.

← Newer: Rent or Buy? What Running the Big Three Actually Costs — Tokens, Hours, or Silicon Older: What Quantization Costs You: the IQ1→Q8 Ladder With Measured Numbers →