Estimate · not a bench

Quant + Hardware Calculator

Three views, one engine: size a model, ask what a Radeon can hold, or read a decode-speed floor. Totals are estimates unless a tested badge shows a first-party R9700 number. Nothing below Q4. llama.cpp first. Qwen2.5 and Qwen3-8B are not offered.

Benchmarks Recommended stacks Quant explainer GPU directory

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.

Radeon

PresetVRAM per GPU (GB)GPU countSystem RAM (GB)Published bandwidth (GB/s)

Split the model across these cardsOff = one full copy must fit a single GPU (house rule: one LLM per GPU). On = weights + KV divided by GPU count, plus extra headroom for PCIe sync (estimate).Keep 5% desktop spareA “fits” answer leaves room for a display. Tight means it only clears with no spare.

ModelPreset

Or read a local GGUF headerHeader only, in this browser. Nothing uploads. Weight size is the file on disk.

Total params (billions)Active params (MoE, billions)Hidden layersKV-growing layersKV headsHead dimKV kindMLA dimSliding-window layersWindow tokens

Quant, context, backend

Nothing below Q4 is listed. Q1 / Q2 / Q3 / IQ1–IQ3 stay off this page, including draft heads.

QuantContext length

KV cache type

Inference backendllama.cpp pathConcurrent sequencesOllama multiplies KV per request. llama.cpp shares one pool.Controller efficiencySpeed-floor only. Default 0.75 is an estimate, not measured.

Host token embeddings / n-gram table in system RAMllama.cpp can keep an untied embed table off the card. Subtracts a formula size — not a measured split.Override overhead (GiB)Framework overhead (GiB)Used only when override is on. Backend floors are estimates, not R9700 measurements.

Required VRAM · estimate

Fits in VRAM (estimate)

Per GPU vs the VRAM you entered. Totals in GiB (1024³). Desktop spare is extra.

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.

Weights: —
KV cache: —
Overhead: —
Total: —

Per GPU · — GiB vs — GB

Usable pool · — GB

Overflow to system RAM · — GiB (—% of RAM)

Best quant that fits fully in VRAM

—

Max context at that quant

—

Max context at the selected quant

—

VRAM vs context

Same estimate at each preset window. Heights are relative to the largest total in the sweep.

Matching Radeon options

Smallest published VRAM first. Fit uses this model, quant, context, and backend — plus desktop spare.

Save this setup

How this result is calculated

weights_GiB ≈ total_params × 10⁹ × bpw / 8 / 1024³A Hugging Face listing or a local GGUF file size replaces the formula when present. Active experts do not shrink the file.

KV_GiB ≈ 2 × full_layers × heads × dim × context × bytes / 1024³plus sliding-window layers at min(context, window), or MLA as layers × dim × tensors × context × bytes. Hybrid Qwen cards only grow token-length KV on full-attention layers.

Backend overhead is a labeled floor plus a percent of weights+KV. llama.cpp Vulkan is the default path; HIP adds a small estimate. These are not R9700 measurements. Ollama multiplies KV by concurrent sequences; the other backends share one pool.

A card fits when the total plus 5% spare lands in the usable pool (more spare on a split).Tight clears the label but not the spare.No overflows — the extra is system RAM offload.

Speed floor: published bandwidth × 75% controller / (active params × bpw / 8). Offered only when the file stays in VRAM. First-party tok/s appear only on the tested badge.

Check price · no live prices

Cards in this VRAM neighborhood

Estimate only. We show the selected card and any listed GPU whose published VRAM is at least the per-GPU total. Buttons open a search — never a stale price. As an Amazon Associate I earn from qualifying purchases.

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.

Quant + Hardware Calculator

What do you want first?

Runs on this pool

Out of reach

VRAM vs context

Matching Radeon options

Cards in this VRAM neighborhood

AMD Radeon AI PRO R9700

AMD Radeon RX 9070 XT

AMD Radeon RX 9070

AMD Radeon RX 7900 XTX

AMD Radeon RX 7900 XT

AMD Radeon RX 7800 XT