24 GB GDDR6X Ampere card. Q4_K_M weights are ~17 GB; measured KV at 64k needs about 22 GB on card. Confirm seller and that it is a 3090 24 GB, not a 3080.
Who it is for
- Local Qwen3.8 27B Q4 on CUDA at 4k–64k context without paying 5090 money.
- Builders who already have a desktop PSU and case that can hold a 350 W-class 3090.
Who should skip it
- Anyone who needs 128k fully GPU-resident context — that wants 32 GB VRAM.
- Quiet deskside DGX OS or Apple-silicon-only shops.
What to compare
- Qwen3.8 27B Q4_K_M is ~17 GB on disk. A 24 GB card is the practical CUDA floor for Q4 plus KV, not a 16 GB card.
- Independent llama.cpp measurements put ~22 GB used at 64k context and ~26 GB at 128k — 128k will not stay fully on this card.
- This is Ampere, not Blackwell. NVFP4 / 50-series kernels are the wrong expectation.
Verify on the Amazon listing
- Confirm 24 GB RTX 3090, not 3080 Ti / 4070 Ti 16 GB lookalikes.
- Used and renewed listings are common. Confirm seller, condition, and that fans/BIOS are intact.
Honest limitation
24 GB is tight past ~64k context. It will not run BF16 27B (~56 GB) or Qwen3.8 Max.
Related picks in this catalog
Affiliate disclosure: As an Amazon Associate I earn from qualifying purchases.
Verify live price
Check price on Amazon
Paid link, placed here for readers who compared the notes first.