How Much RAM: 64 vs 96 vs 128 vs 192GB
Pick a soldered RAM tier by model class (70B, 120B MoE, or 235B Q4), then filter today’s dated price table.
64GB fits about 70B Q4; 96GB is comfortable for 70B but not 120B MoE in the GPU share; 128GB reaches 120B MoE; 192GB PRO 495 reaches 235B Q4. RAM is soldered — buy the tier you need.
Key takeaways
- Unified RAM ≠ GPU-addressable RAM — use the vendor GPU figure when present.
- 64→70B Q4; 96→70B only; 128→120B MoE; 192 PRO 495→235B Q4.
- Memory is soldered; Windows ~96GB GPU caps vs Linux GTT tuning matter.
Who this is for
- Helpful when you are choosing a soldered RAM tier for a target model class (70B, 120B MoE, or 235B).
- Less useful when you plan to upgrade memory later or need measured tokens/sec for a specific runtime.
RAM on these machines is soldered. You cannot add modules later, so the tier you buy is the tier you keep. Figures below are from the price table; model-class labels are the same ones shown on each product row.
Unified memory vs GPU-addressable memory
Unified RAM is the full soldered pool (64 / 96 / 128 / 192GB on this table).
GPU-addressable is the portion the iGPU can use for weights — shown when the vendor publishes it:
- 64GB unified → 48GB GPU-addressable (Framework Desktop 64GB, EVO-X2 64GB, MS-S1 Max 64GB).
- 128GB unified → 96GB GPU-addressable (EVO-X2 128GB, Bosgame M5 128GB, and other 395 128GB rows).
- 192GB unified (PRO 495) → 160GB GPU-addressable (Framework Desktop PRO 495, Bosgame M5 Max, EVO-X5 Pro, MS-S1 MAX-P495).
The Bosgame M5 96GB row leaves GPU-addressable blank on the table. The model-class label on that row is still 70B Q4.
Windows vs Linux: Windows commonly caps GPU-visible memory around 96GB unless you change allocation; Linux can raise GTT / similar limits so more of a 128GB or 192GB pool is usable for inference. Plan the OS with the tier, not after unboxing.
64GB — about 70B Q4
GPU share ~48GB. Model class on these rows: 70B Q4. Do not expect 120B MoE in the GPU share.
| Model | Primary | $/GB | Checked | Stock (primary) |
|---|---|---|---|---|
| $1,959 · Official | $30.61 | October 31, 2026 | Out of stock pre-order | |
GMKtec EVO-X2 64GB · 1TB | $2,200 · Amazon | $34.37 | November 13, 2026 | In Stock |
Minisforum MS-S1 Max 64GB · 2TB | $2,599 · Amazon | $40.61 | November 13, 2026 | In Stock |
96GB — 70B comfortable; 120B MoE does not fit the GPU share
One 96GB row on the table: Bosgame M5 96GB · 2TB.
- Primary Official $2,399 · $/GB $24.99 (checked November 13, 2026).
- Stock: Pre-sale 20–35 days.
- Max model class: 70B Q4 — 120B MoE does not fit the GPU share at this tier.
128GB — about 96GB GPU; 120B MoE fits
GPU-addressable 96GB on published 395 rows. Max model class: 120B MoE.
| Model | Primary | $/GB | Checked |
|---|---|---|---|
| $3,449 · Official | $26.95 | October 31, 2026 | |
GMKtec EVO-X2 128GB · 1TB | $3,500 · Amazon | $27.34 | November 13, 2026 |
GMKtec EVO-X2 128GB · 2TB | $3,650 · Amazon | $28.52 | November 13, 2026 |
Beelink GTR9 Pro 128GB · 2TB | $3,699 · Amazon | $28.90 | November 13, 2026 |
| $3,699 · Amazon | $28.90 | November 13, 2026 | |
ACEMAGIC M1A PRO+ 128GB · 2TB | $3,800 · Amazon | $29.69 | November 13, 2026 |
Minisforum MS-S1 Max 128GB · 2TB | $3,959 · Amazon | $30.93 | November 13, 2026 |
192GB PRO 495 — about 160GB GPU; 235B Q4
Chip: Ryzen AI Max PRO 495. Bandwidth 273 GB/s. GPU-addressable 160GB. Max model class: 235B Q4.
| Model | Primary | $/GB | Checked | Stock (primary) |
|---|---|---|---|---|
| $6,199 · Official | $32.29 | November 13, 2026 | Pre-order first batch | |
| $6,799 · Official | $35.41 | October 31, 2026 | Out of stock; batch 3 Dec 2026 | |
GMKtec EVO-X5 Pro 192GB · 2TB | $6,799 · Official | $35.41 | November 2, 2026 | In stock |
Minisforum MS-S1 MAX-P495 192GB · 2TB | $7,399 · Official | $38.54 | October 31, 2026 | Sold out |
Bottom line
Buy the RAM tier for the model class you need — 70B, 120B MoE, or 235B — because soldered memory cannot be upgraded. Filter the price table by RAM tier to see today’s dated rows.
FAQ
Can I upgrade RAM later?
No. These Strix Halo / Ryzen AI Max configurations use soldered unified memory. Pick the tier at purchase.
Is 96GB enough for 120B MoE?
Not under the 120B MoE label used here. The Bosgame M5 96GB row is labeled 70B Q4. Use a 128GB row labeled 120B MoE when you need that class.
Why does Windows matter for 128GB and 192GB?
Windows often defaults near a 96GB GPU-visible cap; Linux can raise GTT-style limits. Unified capacity is not the same as GPU-usable capacity until the OS allows it.



