Local model finder.

Which local AI model can your computer run?

This free finder compares downloadable GGUF language models against the memory you enter. Choose your computer, set a context length, then compare estimated fit, published scores and exact files. It calculates a shortlist in your browser; it does not run a model or measure your machine. Start small. Your laptop does not need to audition for a data centre.

Hardware: RAM, GPU memory and DGX Spark

Apple silicon Macs use one shared memory pool. Windows and Linux computers can use system RAM for CPU inference, or a dedicated GPU memory budget. The Intel Mac option uses system RAM. With a dedicated GPU, RAM and VRAM are checked separately, never added together. A memory match still needs a runtime that supports your GPU and the model architecture. This is not a multi-GPU calculator.

For ordinary computers, the calculator treats the entered capacity as GiB: it reserves the larger of 3 GiB or 20% of system RAM, and the larger of 0.75 GiB or 10% of dedicated VRAM. The single DGX Spark preset instead uses a 120 decimal GB working budget, about 111.8 GiB, out of its 128 GB unified pool. That reserve is an assumption, not a measurement of your background apps. See the Spark preset and NVIDIA runtime guide.

GGUF, MLX and runtimes

GGUF packages model weights and metadata. The listed files target compatible versions of llama.cpp, Ollama or LM Studio; support varies by architecture and release. Open a model's download panel for its file, model card and runtime notes. An available download is not proof that every runtime can load it. Our Ollama and Python tutorial walks through a first local chat.

MLX LM is another route for local inference on Apple silicon, using its own compatible or converted weights. This directory does not list MLX files or estimate MLX memory. A GGUF fit estimate cannot be transferred directly to MLX. Where a download panel offers vLLM guidance, that refers to the original checkpoint and its runtime requirements, not a promise that the GGUF memory estimate applies.

How memory compatibility is calculated

Estimated memory = GGUF file weights + text attention cache and recurrent state + runtime allowance. The allowance is the larger of 15% of the weights or 0.75 GiB. File sizes use decimal GB; ordinary memory results use GiB. The text cache grows with your selected context length, with separate sliding-window limits and fixed recurrent state where recorded. The cache assumptions use FP16; changing cache precision or runtime settings changes real usage.

Comfortable leaves at least 12% of the usable budget free. Tight fit fits with less headroom. CPU only fits system RAM but exceeds the dedicated GPU budget. Too large exceeds the estimated pool; partial CPU/GPU offload may still be possible but is not estimated. Context limit means the requested context exceeds the catalogued model limit. Spark estimates must stay strictly below 120 GB.

These estimates cover one text session. Concurrent requests, image projectors, audio processing, speculative draft models and extra runtime allocations can need more memory. A model fitting on paper does not establish its speed, quality or ability to load on your setup. Check the model card, close other apps and test a short prompt before increasing context.

Quantization and ranking: what the numbers mean

Quantization reduces weight precision and usually download size, with possible quality changes. Q4 and Q8 labels describe weight formats, not a guarantee of equal accuracy or speed across models. Auto chooses an available Q8_0 file when it leaves at least 25% of the usable budget free; otherwise it chooses the catalog's preferred file in the best fit tier. The estimates use the selected file size, not parameter count alone.

The directory starts with the newest catalogued releases that fit your filters. Sorting and benchmark selection are separate: changing a test keeps the release order unless you choose Highest selected score. Exact release days appear where verified; entries with only a known month follow dated releases within that month. Repository upload or edit dates are not automatically treated as model release dates.

Five capability groups make the scores easier to explore: reasoning (GPQA Diamond and AIME), coding (LiveCodeBench and SWE-bench), agents and tools (BFCL and Terminal-Bench), instructions (IFBench and IFEval), and conversation (MultiChallenge). These tests measure different things. No benchmark is reported by every publisher, and versions, subsets, precision and evaluation harnesses can change the result. Each comparison keeps its evaluator and test conditions attached. General indexes and specialist tests remain under More tests.

Each mini chart shows two scores on the same 0-100 scale: the local model and GPT-5.1, or another advanced model when a GPT-5.1 result is unavailable. The default GPQA Diamond comparison uses Artificial Analysis results with GPT-5.1 at high reasoning effort. A benchmark score describes performance on that test, not overall intelligence. Reference models run in the cloud and are excluded from memory estimates and downloads. Hover over a label for the tested configuration, or follow a score to its source.

Sorting by benchmark ranks only results from the selected test and evaluator. Models with other published results appear in separate groups, each ordered on its own test. Use Rank this test to move one of those evaluations into the main ranking. Compatible models come first within each group. Different tests are never averaged together. When the comparison bars use a different test from the ranked score, the score used for sorting appears prominently above them. Publisher results can appear beside an independent reference on the same benchmark, labelled as different test setups. Those contextual comparisons receive no calculated gap because prompts, precision, agent scaffolds and scoring setups can differ. Missing results are never displayed as zero. Open a model's published-results link to see every evaluation and its test conditions. Ties share the top local score; smaller working sets come first within a tie. The hardware-fit sort groups fit tiers, then uses a parameter-size heuristic; it does not measure inference speed.

Follow each score's evidence link for its source and setup. Published scores may use different precision and settings from your selected file. No laptop speed benchmarks are implied. For an example of reporting actual hardware, settings and limitations, try our reproducible thinking-mode experiment.