The Hardware card in Settings

Short answer

Settings → Hardware shows what Abliteration sees on your machine - CPU, total RAM, GPU, and free RAM right now - and which model sizes fit comfortably. It’s the same info as the first-launch popup, always available.

Detailed explanation

Abliteration reads your hardware through standard Windows APIs on startup and caches the result. The Hardware card displays three or four lines depending on your GPU:

Line 1 - Your PC. A one-line summary: Your PC: {CPU name}, {N} GB RAM, {GPU name}. The GPU name is the actual card the machine has - Intel Arc Graphics, NVIDIA GeForce RTX 4070 (12 GB VRAM), AMD Radeon RX 6600, whatever Windows reports. Trademark marks like (R) and (TM) are stripped, but the rest of the name is kept verbatim because it identifies the specific part. Green when the card can accelerate inference (modern NVIDIA), amber otherwise. When Windows cannot enumerate any real GPU, this reads No GPU detected.

Line 2 - Free right now. Free right now: {X.X} GB. What’s available at this moment, after Windows and your other apps. This is the number that determines whether a model will load smoothly right now, more than the total.

Line 3 - What you can run. Compact summary like You can run: 3B (now), 7B (close apps), 13B (close apps), 30B (close apps). 70B won't fit. Colors flag the tiers: green (fits in free RAM), amber (needs closing apps but fits in total RAM with headroom), red (won’t fit at all).

Line 4 - GPU acceleration status (only when acceleration isn’t available). One muted-grey paragraph:

GPU acceleration is not available on this PC. Abliteration runs on your CPU.

This is the same sentence for every non-usable case - integrated iGPU, older NVIDIA, AMD Radeon RX, Intel Arc discrete, virtualised guest, no GPU found - and it’s absent on machines with a modern NVIDIA card the app CAN accelerate on. See Hardware detection for why the same sentence covers every non-usable case.

The Hardware card does NOT display CPU cores/threads or disk free space, even though those values are probed. They’re used internally (for tok/s estimates) but not shown in the UI.

What DOESN’T appear. A DisplayLink USB dock, a Parsec virtual display, a remote-desktop adapter, Windows’ Basic Display Adapter - none of these show up on Line 1 even though Device Manager lists them. They’re display sinks, not compute devices. See Hardware detection for the filter rule (one line: the device must be on the PCI bus).

Why it matters

Before adding a big model, glance at Hardware to see if you’re realistically going to run it. If Line 3 says “30B (close apps)” and you want 30B - close Chrome first. If it says “won’t fit” - either drop to a smaller size, use a heavier quantization (Q4 instead of Q8), or don’t bother.

In Abliteration specifically

Path: Settings → Hardware. Also shown as a popup on first launch (dismissible with “Understand” - never shown again after that).

The “You can run” line uses calibrated size estimates for each tier: 3B ≈ 3.5 GB, 7B ≈ 7 GB, 13B ≈ 12 GB, 30B ≈ 22 GB, 70B ≈ 50 GB. These include model weights + KV cache at 8K context + Windows overhead. Real models vary within a tier depending on quantization. The amber tier requires ~6 GB headroom above the need.

GPU is detected but not used for inference - CPU-only. The tok/s tooltip on model rows may name your NVIDIA GPU in the “on your {hardware}” phrasing, but the estimate itself is a CPU one (this is a small labeling quirk to be aware of, not a functional issue).

Related

• Hardware detection - how the detection actually works, and why some devices in Device Manager never appear here
• Understanding your rig - first-launch tour of these numbers
• RAM requirements - what “13B needs 12 GB” actually means
• Fit-check dialog - the specific warning when you add an oversized model