Guides
Used GPU keep or sell: VRAM 6GB is the dividing line
Strata gives used GPUs three more years. Decision rule: VRAM ≥ 6GB keep, < 6GB sell. RTX 3050 4GB cannot fit 27B quantizations.
Used GPU keep or sell: VRAM 6GB is the dividing line
Strata author Niko one week ago: 6GB VRAM + three-tier offload + MTP = 5090 performance. This is the decision.
Accounting (for 27B Q4_K_M ≈ 16GB model)
| Card | VRAM | Now price | Keep / Sell | Why |
|---|---|---|---|---|
| GTX 1080 | 6GB | ¥600-800 | ✅ Keep | Strata three-tier offload + MTP can carry |
| GTX 1080 Ti | 11GB | ¥1100-1300 | ✅ Keep | Enough VRAM for IQ2_XS 27B |
| RTX 2060 | 6GB | ¥700-900 | ✅ Keep | Same as 1080 |
| RTX 3050 | 4GB | ¥500-600 | ❌ Sell | 4GB can't even hold IQ2_XS |
| RTX 3060 | 12GB | ¥1400-1700 | ✅ Keep | Sweet spot transition card |
| RTX 4060 | 8GB | ¥1600-1800 | ✅ Keep | Strata MTP 1.68× speedup |
| RTX 5070 | 12GB | ¥3500-4000 | ✅ Keep | 27B IQ2_XS = 79 tok/s measured |
| RTX 5090 | 32GB | ¥16000+ | ✅ Keep | Strata on 32GB ≈ "overkill but faster" |
Keep vs sell
✅ Keep: VRAM ≥ 6GB + Strata compatible + not cash-strapped ❌ Sell: VRAM < 6GB (can't run 27B quant) OR need cash OR upgrading to 32GB+ long-term
Three judgment points
1️⃣ Premise: Niko's "6GB = 5090" is a forward-looking claim, not currently reproducible 2️⃣ Strata v0.5.x is transitional. Qwen 4 full support = v0.6 (no clear timeline) 3️⃣ Even with v0.6, HDD will kill SSD offload — >3 GB/s sequential read NVMe is the floor
Evidence
- "6GB = 5090": Niko Veit @coldniko 2026-10-06 (forward-looking)
- 79 tok/s @ 5070 12GB: local DB Strata record id 52
- 70 tok/s @ 3060 12GB: local DB Strata record id 51
- Prices: Xianyu 2026-10 samples (non-merchant, reference only)
Sources
- Niko tweet: https://x.com/coldniko/status/2107406312682258511
- Local DB Strata records (id 51/52/53)
- Prices: Xianyu 2026-10 samples