About Local Model Guide
Local Model Guide — localmodelguide.com — answers one question: can I run this model on this hardware, and how fast?
Running local models used to mean digging through forum posts, cross-referencing spec sheets, and guessing at VRAM requirements. Local Model Guide turns that into a simple lookup. We track open-weight LLMs, the machines people actually run them on, and the quantization formats that decide whether a model fits, then we put the three together on one page so you can see at a glance what will work on your hardware.
What the site covers
Every model page shows how much memory the model needs at each quantization, which popular GPUs and Macs can hold it, and any measured tokens/sec results the community has published for it. Every hardware page works the other way around: the largest models that device can run, the best-fitting quantization for each, and its price per GB of memory. The Quickstart grid condenses all of it into one popular-models × popular-hardware matrix, and the guides explain the concepts behind the numbers in plain language.
Who it's for
Developers picking a model to run against a local API, hobbyists squeezing the most out of the machine they already own, and buyers deciding between a high-memory Mac and a dedicated GPU. If you know what a GPU is but not what GGUF means, you're the audience: we define each term once and then get out of the way. No jargon without explanation, no marketing fluff, just the numbers you need to make a decision.
How the numbers work
Memory fit is estimated, and we say so wherever it appears. Weights are calculated from the parameter count and the quantization's bits per weight, plus headroom for the KV cache at a stated context length. That estimate is compared against usable memory: about 90% of VRAM on a discrete GPU, about 75% of unified memory on Apple Silicon. The result is one of three verdicts: runs, partial offload, or does not fit. We show the formula on the methodology page so you can verify the math yourself.
Throughput is different: tokens/sec figures are measured, not estimated. They come from community benchmarks, and they are always labeled as measured. Runtime, context length, and settings all move the real number, so treat them as a guide rather than a guarantee. We attribute each benchmark to its source and show the date it was recorded.
What we don't do
We don't rank models on quality, and we don't sell hardware. Local Model Guide is a neutral reference, not a retailer and not a leaderboard. If we ever add affiliate links, we'll disclose them clearly, and they won't change a verdict or a ranking. Our only goal is to help you make an informed decision about what to run and what to buy.