Buy the box

Setup by budget

Four tiers from a 16 GB card to the two-R9700 test rig. Answer size vs speed first — that changes the pick. Models and quants are Q4 and up, estimated through thecalculator. Backend default is llama.cpp. Buttons say “Check price” — we do not show a live or stale price.

Price bands as of 2026-10-12: approximate, check current price — not a live or stale quote.Every USD band is still TODO — no cited street price, so the slider is by tier name, not dollars.

Ask first

What do you want first?

Three questions before the recommendation. They change the pick, not the VRAM math. Tokens per second are estimates unless a “Tested on our rig” number is showing. No cloud prices.

Priority

Max model sizeMore VRAM, more RAM, Q5/Q6 when they fit. Slower tokens.SpeedOne GPU, no offload, MoE, Q4_K_M to Q5.

Usage

Fully localBigger VRAM/RAM and long context. The local model does everything.Hybrid (local plus a cloud model like Claude, Grok or ChatGPT for token-heavy work)Local keeps privacy-sensitive, repetitive, or always-on work. Cloud can take long jobs. No cloud prices on this page.

Model style

Standard (original)Official instruct. The default.Abliterated / uncensoredEducational. Can hurt quality. You are responsible for use.

Abliterated hub— what it is, how to vet files, the catalog. Educational only. Tested pick: being reworked.

Which tier are you in?

Starter
Mid
Enthusiast
Pro

Optional dollar input

No cited USD bands — pick a tier on the slider. Fill usd_low / usd_high in budget-tiers.yaml when you have a source.

Save this setup

Abliterated / uncensored

Abliteration edits refusals in the weights. It can hurt quality. Verify the source and the uploader. You are responsible for what you run. Floor is still Q4 — nothing below. Read theabliterated hub.

Tested on our rig: being reworked. Do not invent a daily-driver swap.

Huihui Qwen3.8-27B abliterated
huihui-ai/Huihui-Qwen3.8-27B-abliterated
Classic abliteration (remove-refusals-with-transformers style). Community GGUF / Ollama tags exist — confirm the exact file. Same 27B envelope if you pick a 4–5 bit quant. Confirm the file. Floor is still Q4.
Primary educational row this pass. Publisher frames it as experimental. Some builds keep early layers unablated. Test think-loops. · llama.cpp first. MTP/vision "not modified" on the card. Pin a revision. Not a Stack Lab profile. Not the daily driver.
Huihui Qwen3.5-9B abliterated
huihui-ai/Huihui-Qwen3.5-9B-abliterated
Classic abliteration (remove-refusals-with-transformers) of Qwen/Qwen3.5-9B. llama.cpp GGUF via mradermacher/Huihui-Qwen3.5-9B-abliterated-GGUF (Q4_K_M ~5.7 GB, Q8_0 ~9.6 GB, f16 ~18.0 GB on that card). 4–8 bit GGUF should leave room on 32 GB if the file matches. f16 is tighter once cache grows. Confirm.
Smaller educational row. Unverified on our rig. Qwen3.5 may need a current llama.cpp. · llama.cpp first. Publisher research / experimental framing. mmproj files exist on the GGUF card — confirm whether you want vision tensors. GGUF sibling mradermacher/Huihui-Qwen3.5-9B-abliterated-GGUF.
orcarouter Qwen3.8-27B Uncensored
orcarouter/Qwen3.8-27B-Uncensored
Abliteration / refusal-direction work on Qwen3.8-27B. Q4_K_M ~16.8 GB (publisher figure). Fits with context room if the card is honest. Still never below Q4.
Community writeups exist; still verify MTP/vision claims on your build · llama.cpp first. Some GGUF repos are gated. Read the license and safety note on the card. GGUF sibling published — confirm which file.
Heretic-style uncensored GGUF (e.g. JonathanColetti)
JonathanColetti/Qwen3.8-27B-Uncensored-GGUF
Heretic refusal-direction minimization + MTP graft notes on the card. Q4_K_M and others listed on the card. Treat Q4 as the 32 GB daily quant until you measure cache
Card is unusually explicit about what quantization does to the refusal boundary · llama.cpp first. Lower bits can change refusals again. Evaluate behavior on a high quant first.
SFT-class refusal removal (e.g. Abliterated-SFT)
jenerallee78/Qwen3.8-27B-Abliterated-SFT
Small SFT / LoRA on refusal, not only direction ablation. Publisher GGUF notes exist — read GGUFs.md. Unknown until you pick a quant. Do not invent a fit.
Different failure mode (harness/system-prompt can restore refusals) · llama.cpp first. Finicky with agent frames. Useful as a contrast, not the default.

Catalog as of 2026-10-12. Fit notes are the catalog’s 32 GB column, not a new calculator run — we do not invent KV math for these files. Q4 and up only. Confirm the live model card.

One ~16 GB Radeon
Starter
Suggested for this priority
~16 GB (RX 9070 XT or RX 7800 XT)
TODO: no cited street-price source as of 2026-10-12
RAM
64 GB DDR5 is the one-card minimum this site already uses as search guidance — not a measured requirement.
NVMe
House rule — GGUFs on NVMe, not a spinning disk. Search only, not a SKU.
PSU
TODO: no cited PSU wattage for the 16 GB cards on this page. AMD’s R9700 minimum is 750 W (product page). Check the card you buy.
Backend
llama.cpp (Vulkan first; HIP if rocminfo is clean). Not Ollama, not vLLM. See the chooser.
Speed wants the whole file in one GPU with no offload. The house Q4+ GGUFs do not fit 16 GB. This tier cannot be a speed pick for those models. Do not drop below Q4 to fake a fit.
What fits (Q4 and up)
- Qwen3.6-35B-A3B · Q4_K_M
  Estimate only. Q4 formula weights for 35B already exceed 16 GB, so expect RAM offload (slow). Do not drop below Q4.
  Open this setup in the calculator
- Qwen 3.8 Flash · UD-Q6_K_XL
  Publisher file ~169 GB. Will overflow 16 GB by a lot. Shown so the calculator can say no. Not a speed pick.
  Open this setup in the calculator
What you give up. Versus Mid — a 20–24 GB card that may hold Q4 at a modest context with no offload (estimate).
Check price · no live or stale prices
AMD Radeon RX 9070 XT
Check price · no live or stale prices
AMD Radeon RX 7800 XT
Check price · no live or stale prices
High-capacity DDR5 kit
Search only — not a specific SKU. Our test rig is 256 GB DDR5; 64 GB is a reasonable one-card minimum.
Check price · no live or stale prices
Fast NVMe SSD
House rule — load GGUFs from NVMe, not a spinning disk. Search only, not a SKU.
Check price · no live or stale prices
1000W-class PSU
AMD’s R9700 minimum recommendation is 750 W. A 1000W-class unit is search guidance for headroom, not a SKU.
One ~20–24 GB Radeon
Mid
Suggested for this priority
~20–24 GB (RX 7900 XT or RX 7900 XTX)
TODO: no cited street-price source as of 2026-10-12
RAM
64 GB DDR5 search guidance. More RAM only helps if you offload — and offload is slow.
NVMe
Same house rule — NVMe, not HDD.
PSU
TODO: no cited PSU wattage for the 7900-class cards on this page. Check the product page for the SKU you buy.
Backend
llama.cpp (HIP gfx1100 / gfx1101, or Vulkan). Still not vLLM. See the chooser.
One 24 GB card, Q4_K_M, shorter context, no split. Q5 at long context is a size move and likely overflows. Offload kills speed. Tokens per second — estimate only; no first-party number on a 7900.
What fits (Q4 and up)
- Qwen3.6-35B-A3B · Q4_K_M
  On 24 GB, Q4 + a modest context can look like a VRAM fit in the calculator. On 20 GB it is tight or no. Open the share link — do not take this sentence as a bench.
  Open this setup in the calculator
- Qwen3.6-35B-A3B · UD-Q5_K_XL
  Publisher 26.6 GB file plus KV will not sit in 24 GB at 262K. Try a shorter context in the calculator before you leave Q5. Size path — offload is slow.
  Open this setup in the calculator
What you give up. Versus Enthusiast — a 32 GB R9700 that holds the daily-driver Q5 at 262K with 0% offload (tested on our rig). Measured on our rig, 2026-10 — UD-Q5_K_XL ~50 tok/s at 256K (100.8 short, ~88 at 32K).
Check price · no live or stale prices
AMD Radeon RX 7900 XTX
Check price · no live or stale prices
AMD Radeon RX 7900 XT
Check price · no live or stale prices
High-capacity DDR5 kit
Search only — not a specific SKU. Our test rig is 256 GB DDR5; 64 GB is a reasonable one-card minimum.
Check price · no live or stale prices
Fast NVMe SSD
House rule — load GGUFs from NVMe, not a spinning disk. Search only, not a SKU.
Check price · no live or stale prices
1000W-class PSU
AMD’s R9700 minimum recommendation is 750 W. A 1000W-class unit is search guidance for headroom, not a SKU.
One 32 GB R9700
Enthusiast
Suggested for this priority
32 GB (one Radeon AI PRO R9700)
TODO: no cited street-price source as of 2026-10-12
RAM
64 GB is the one-card search floor. Our daily driver used 0% RAM offload; extra DRAM is still useful for the OS and for the Flash file if you try it.
NVMe
NVMe. Daily-driver GGUF is a 26.6 GB listing — do not park it on a spinner.
PSU
AMD product page minimum for the R9700 is 750 W. 1000W-class is search guidance for headroom, not a SKU.
Backend
llama.cpp Vulkan (the daily-driver path on our rig). HIP if you already have a clean /opt/rocm. See the chooser.
This is the speed tier — Qwen3.6-35B-A3B Q5 on one R9700, no offload (tested on our rig). Measured on our rig, 2026-10 — UD-Q5_K_XL 100.8 tok/s short, ~88 at 32K, ~50 at 256K. Flash on one card overflows; it is not a speed pick.
What fits (Q4 and up)
- Qwen3.6-35B-A3B · UD-Q5_K_XL
  Tested on our rig — Qwen3.6-35B-A3B Q5
  Daily driver on our rig — one R9700, 262144 context, f16 KV, 0% RAM offload. Measured on our rig, 2026-10 — UD-Q5_K_XL ~50 tok/s at 256K (100.8 short, ~88 at 32K). Exact flags / UD-Q5_K_XL VRAM still TODO.
  Open this setup in the calculator
- Qwen 3.8 Flash · UD-Q6_K_XL
  Publisher ~169 GB. One 32 GB card will overflow; the calculator says so. Pro is where we actually ran it (RAM holds most of the file). Not a speed pick.
  Open this setup in the calculator
What you give up. Versus Pro — extra DRAM and a second card help the size path, not this one-card speed shape.
Check price · no live or stale prices
AMD Radeon AI PRO R9700
Check price · no live or stale prices
High-capacity DDR5 kit
Search only — not a specific SKU. Our test rig is 256 GB DDR5; 64 GB is a reasonable one-card minimum.
Check price · no live or stale prices
Fast NVMe SSD
House rule — load GGUFs from NVMe, not a spinning disk. Search only, not a SKU.
Check price · no live or stale prices
1000W-class PSU
AMD’s R9700 minimum recommendation is 750 W. A 1000W-class unit is search guidance for headroom, not a SKU.
Two R9700 + 128–256 GB DDR5
Pro
Tested on our rigSuggested for this priority
2 × 32 GB (our test rig). Daily driver still uses one card.
TODO: no cited street-price source as of 2026-10-12
RAM
128–256 GB DDR5. Our test rig is 256 GB. The Flash UD-Q6_K_XL listing is ~169 GB — that is why Pro has this much DRAM.
NVMe
Fast NVMe. Two large GGUFs plus the OS.
PSU
Two 300 W TBP cards (AMD R9700 TBP, product page) plus a 750 W minimum per card is not a PSU math we will invent. 1000W-class is search guidance. TODO — cite a PSU pick once we have one.
Backend
llama.cpp Vulkan for the daily driver. Strata v0.1.42 for the Flash agent. vLLM remains a try. See the chooser.
Daily driver still uses one of the two cards. Qwen3.6-35B-A3B Q5, 0% offload (tested on our rig). Do not split for speed. Flash is the size pick. Measured on our rig, 2026-10 — UD-Q5_K_XL 100.8 / ~88 / ~50 tok/s.
What fits (Q4 and up)
- Qwen3.6-35B-A3B · UD-Q5_K_XL
  Tested on our rig — Qwen3.6-35B-A3B Q5
  Tested on our rig — one of the two cards, llama.cpp Vulkan, 262K, f16 KV, 0% offload. Measured on our rig, 2026-10 — UD-Q5_K_XL ~50 tok/s at 256K (100.8 short, ~88 at 32K).
  Open this setup in the calculator
- Qwen 3.8 Flash · UD-Q6_K_XL
  Tested on our rig — Qwen 3.8 Flash Q6
  Tested on our rig — Strata v0.1.42 ~32.6 tok/s vs llama.cpp ~16. Publisher file ~169 GB; VRAM/RAM figures still TODO. Size pick, not the speed pick.
  Open this setup in the calculator
What you give up. Nothing on the speed ladder — this already matches the one-card daily driver. Flash Q6 is the size move.
Check price · no live or stale prices
AMD Radeon AI PRO R9700
Check price · no live or stale prices
High-capacity DDR5 kit
Search only — not a specific SKU. Our test rig is 256 GB DDR5; 64 GB is a reasonable one-card minimum.
Check price · no live or stale prices
Fast NVMe SSD
House rule — load GGUFs from NVMe, not a spinning disk. Search only, not a SKU.
Check price · no live or stale prices
1000W-class PSU
AMD’s R9700 minimum recommendation is 750 W. A 1000W-class unit is search guidance for headroom, not a SKU.

As an Amazon Associate I earn from qualifying purchases. Some Newegg links use Rakuten Advertising and some B&H links use B&H’s affiliate program when those IDs are configured. Micro Center and AMD Store links are plain “also available at” links — no tracking. Buttons say “Check price” and never show a live or stale price.Full disclosure.

Already built