Recommended machines
Hearth runs models on your own PC, so the machine decides how big a model feels fast. Here are real, current products, grouped by what they can run.
As an Amazon Associate, Hearth earns from qualifying purchases. Links to Amazon on this page are affiliate links: if you buy, we may earn a commission at no extra cost to you. It does not change what we recommend. Items from other stores are plain links.
Listings were checked on November 12, 2026. Prices and stock change daily and graphics-card and memory prices have been unusually high this year, so we do not print prices: use Check current price. Speeds are rough community measurements with local models, not promises. Hearth is Windows-only today.
Quick guide
Video memory (VRAM) is the number that matters most. As a rule of thumb a 4-bit model needs a little over half a gigabyte per billion parameters, plus room for the conversation. A model that does not fit spills into system RAM and slows down a lot.
| Video memory | Models that fit (4-bit) | What it feels like |
|---|---|---|
| No graphics card, or under 6 GB | Models of about 3 to 4 billion parameters, or larger ones partly in system RAM | Works, but slower. Good for trying Hearth out. |
| 8 GB | 7 to 8B models at 4-bit; 12B only with short context | Hearth’s tested class. Comfortable for everyday chat, summaries and tools. |
| 12 GB | 12 to 14B models at 4-bit | A good middle: noticeably more capable models, still quick. |
| 16 GB | 14B comfortably, and about 20B (gpt-oss-20b is about 14 GB) | The sweet spot for most people who want a bigger model. |
| 32 GB | 27 to 32B at 4-bit, with long context | Enthusiast. 70B-class models still do not fit entirely. |
| 64 to 128 GB unified memory | 70B-class models, and gpt-oss-120b (about 65 GB) | Very large models fit, but dense ones are slow because memory is slower than a graphics card’s. |
Sizes are approximate and vary by model and context length. System RAM matters less than video memory, but 16 GB is a comfortable floor and 32 GB is better. Sources for the speeds quoted below: CanRun (RTX 5060 Ti 16 GB), Database Mart (RTX 5090), Petronella Technology Group (Ryzen AI Max+).
Entry
Everyday 7 to 8 billion parameter models (such as Llama 3.1 8B) at 4-bit, entirely on the graphics card. 12B models fit only with short context. Fast enough that replies stream at a comfortable reading pace.
This is the class of machine Hearth is tested on: an RTX 5060 Ti 8 GB with 16 GB of RAM.
The exact model Hearth is tested on. It is sold by Walmart rather than Amazon, so this link is not an affiliate link.
The same CPU, graphics card, memory and SSD as the model above, in a white case, sold on Amazon.
Sweet spot
Everything in the entry tier with room to spare, plus 12 to 14 billion parameter models comfortably and ones around 20B (gpt-oss-20b is about 14 GB) on the card. Community measurements on an RTX 5060 Ti 16 GB with Ollama report roughly 80 tokens per second for Llama 3.1 8B and about 50 for Gemma 3 12B.
Video memory is the number that matters most. Doubling it from 8 GB to 16 GB opens up a whole class of larger models.
The most direct step up from Hearth’s tested PC: the same graphics-card family with double the video memory and double the RAM.
Same 16 GB graphics card with a larger SSD, which helps because models are several gigabytes each. When we looked it was sold by a third-party seller on Amazon.
The same 16 GB of video memory, but a much faster graphics card, so the same models answer sooner. Worth it if you also game or create.
Enthusiast
27 to 32 billion parameter models at 4-bit entirely on the card, with long context. Community measurements with Ollama report roughly 45 tokens per second on a 32B model. 70B-class models still do not fit and spill into system RAM, which is slow.
A card this size needs a big power supply and case. Buying a complete PC means that part is already sorted.
A ready-built RTX 5090 desktop with 32 GB of video memory, the most you can get on a consumer graphics card.
Mini PC
These share one big pool of fast memory between the processor and the graphics, so very large models fit that no consumer graphics card can hold: 70B-class models, and mixture-of-experts models such as gpt-oss-120b (about 65 GB). A third-party benchmark of a laptop with the same chip family reports roughly 40 tokens per second for gpt-oss-120b on Windows with Ollama. Dense 70B models read their whole weight set for every word, so they are much slower.
We have not tested Hearth on one of these. Ollama’s AMD graphics support on Windows differs by chip and version, and some owners report needing the Vulkan backend, so read Ollama’s GPU notes before you buy.
A tiny desktop with 128 GB of unified memory, sold by the maker on Amazon. Reviews of the EVO-X2 line are mixed on fan noise, so expect to tune its power mode.
The repairable, well-documented alternative, sold directly by Framework. All configurations showed as out of stock when we looked, so check its page.
AI workstation
NVIDIA says one DGX Spark can run inference on models up to about 200 billion parameters, using 128 GB of unified memory at 273 GB/s. It is excellent for big models.
Honest caveat: DGX Spark runs Linux (NVIDIA’s DGX OS), not Windows, and Hearth is a Windows app today, so Hearth does not run on it directly. The way to use one with Hearth is as a model server: run Ollama on the Spark and add it in Hearth under Models, Add remote Ollama host. We have not tested that pairing ourselves. OpenAI-compatible connections are planned for Hearth 0.5.1 to widen what Hearth can talk to.
Not sold on Amazon, so this goes to NVIDIA’s own store (no affiliate tag). It showed as out of stock on the day we checked; 64 GB versions exist from partner brands.
Upgrades
The most effective upgrade for local models is more video memory. Before you buy, check that the card fits your case, that your power supply has the right connectors and wattage, and that you have a free PCIe x16 slot.
The affordable way to move from 8 GB to 16 GB of video memory. A 2.5-slot card with dual BIOS.
Same 16 GB, with a much faster chip on a 256-bit bus, so answers arrive sooner.
The biggest consumer card. A 3.6-slot design, so check your case and power supply first. Sold by a third-party seller when we looked.
Upgrades
System RAM is the cheaper upgrade and helps when a model is a little too big for your graphics card, but it will not make a large model fast. Memory prices have been unusually high in 2026, so compare before you buy. Match the type to your motherboard.
For a modern DDR5 desktop: takes a 16 GB PC to 32 GB.
For an older DDR4 desktop. DDR4 and DDR5 are not interchangeable, so check which your motherboard uses.
Hearth runs on a regular Windows PC. Check the requirements, then download it.
As an Amazon Associate, Hearth earns from qualifying purchases.
Hearth