Best GPU for a Local AI Companion on Windows

There is no single best Windows GPU for every local AI companion. First check whether the app’s model backend supports your exact card and drivers. Then compare memory capacity, expected speed and your budget. VRAM matters, but it is not the only specification that matters.

For Local Waifu, an important correction is needed: a 16 GB GPU does not select Balanced by itself in the Windows source reviewed on October 18, 2026. The selector keeps that value in Light. A machine with more system RAM can still receive a higher recommendation.

This is a source-based explanation of selection and provisioning, not a hardware benchmark. I build Local Waifu, and no tokens-per-second measurements or guaranteed model quality rankings are claimed here.

The actual tier boundaries are consistent across the examples

The short version: The selector takes the higher RAM or detected VRAM tier. It does not combine their capacities.

The inspected Windows implementation uses these bands:

These are application recommendation bands, not universal memory requirements for all language models. The function also returns Light below the supported minimum; that does not make a PC below the published requirements a supported configuration.

For a 16 GB RAM PC with a detected 16 GB NVIDIA card, both sides map to Light. With a detected 24 GB card, the VRAM side maps to Balanced. With 64 GB system RAM and an 8 GB card, the RAM side maps to Power. That last recommendation says nothing about whether all model layers fit on the GPU.

Do not add the two numbers. A 32 GB RAM PC with 24 GB detected VRAM does not become a 56 GB combined pool for this selector. It remains Balanced because neither resource crosses the next boundary.

Capacity and generation speed answer different questions

The short version: Enough memory makes a configuration possible; backend support and hardware performance determine how comfortable it is to use.

Model weights occupy memory, and the working context needs space too. Quantization changes that footprint. Longer context, other applications and the display workload can reduce available headroom even when the card’s advertised capacity looks sufficient.

A larger model may fit partly in system RAM while only some layers are GPU-accelerated. That can work, but response latency can be very different from a model fully loaded onto the GPU. The tier label is not a performance certificate and should not be used to promise a particular conversational pace.

Ollama’s FAQ documents ollama ps, whose processor column distinguishes CPU, GPU and mixed placement. Use that evidence when diagnosing speed instead of assuming that a detected card means every layer is accelerated. In Local Waifu, the hardware settings guide explains the built-in checks.

AMD provisioning exists, but detection is not identical to NVIDIA

The short version: The app has an AMD runtime path, while its separate automatic VRAM tier probe is NVIDIA-only.

The reviewed Windows provisioning code detects AMD and downloads the matching Ollama ROCm runtime. The pinned runtime in this source is v0.30.8. The installer does not already contain every vendor runner. A runtime download needs internet, and failures leave the bundled CPU path available rather than proving GPU acceleration succeeded.

The provisioning code attempts to restart the model engine after a successful download. If applying the runtime fails, a later app restart may be needed. Settings also offers Retry GPU setup. Those are useful recovery paths, not a guarantee that every Radeon card will work without attention.

Separately, Windows hardware detection obtains VRAM through NVIDIA tooling. AMD vendor detection for downloading a runtime does not change that function into an AMD VRAM probe. If no NVIDIA VRAM is detected, the tier selector uses system RAM. The previous claim that all AMD cards automatically receive the same VRAM-based tier as NVIDIA cards was therefore too broad.

Ollama’s current hardware documentation lists Windows AMD support separately from Linux and describes additional Vulkan support. Those current upstream lists are a compatibility reference, not proof that every upstream capability is enabled in Local Waifu’s pinned runtime. Check the exact installed app, driver and backend before purchasing an AMD card for it.

A buying decision should start with the machine you own

The short version: Test the existing configuration before buying more memory or a different GPU.

For ordinary conversation, the relevant question is whether the reply arrives at a pace you enjoy. A smaller model that responds comfortably may suit a quiet evening better than a larger recommendation that spends too long on each answer. No tier label can settle that preference for you.

If the installed model is slow, check placement and driver support first. If it cannot load, inspect memory pressure and context settings. If a particular model is essential, look up that model’s actual file size and runtime requirements rather than buying solely to cross an app’s recommendation boundary.

Used cards can be worth considering, but this guide does not supply a current price comparison or claim that a particular older card is always cheapest. Check the exact SKU, memory capacity, power supply requirements and seller’s return terms. Laptop variants also need their own specifications; a shared product name does not establish desktop-equivalent memory or cooling.

Multiple GPUs require backend-specific checking

The short version: Do not treat a single-card tier probe as a general claim that models cannot span multiple GPUs.

Ollama documents spreading a model across available GPUs when it cannot fit entirely on one. Local Waifu’s inspected Windows recommendation probe, however, uses the maximum detected NVIDIA card capacity rather than summing every card. Model placement and recommendation selection are different layers.

Before buying two cards, verify that the backend supports the intended arrangement and that your machine has room, power and cooling for it. This article makes no universal claim about the largest consumer card or the impossibility of reaching a tier with one GPU.

Sources and a practical next step

The short version: These corrections describe the reviewed code; verify your installed build before relying on its recommendation.

Product-source review: src-tauri/src/inference/hardware_tier.rs, src-tauri/src/platform/windows.rs and src-tauri/src/sidecars/gpu_runtime.rs in the Windows application. The selector and its tests establish the thresholds and maximum-of-resources rule; the platform probe establishes the NVIDIA-only VRAM input; provisioning establishes the AMD download and fallback path.

Public references reviewed October 18, 2026: Ollama hardware support, Ollama FAQ, Local Waifu requirements and hardware controls. This review did not run Windows GPU benchmarks.

For app choices, read the local companion shortlist. For the difference between local text and online extras, use the offline apps guide. A compatible GPU helps the local route; it does not decide whether a separately selected cloud provider receives a prompt.

Questions people ask

Back to the blog