localllm
About this site
localllm is a database that answers "will this local LLM (GGUF quantized model) run on my PC?" using real file sizes and a published formula for the memory it needs.
- Models listed
- 1,276
- GPUs and Macs listed
- 91
- Data last updated
What you can find here
- GGUF file sizes per model and quantization (the actual sizes on Hugging Face)
- Estimated memory (VRAM or unified memory) per context length, and the GPUs and Macs it fits
- Licenses (the original model's terms), whether commercial use is allowed, and the type of distributor
- Popularity by Hugging Face downloads, and third-party evaluations (LMArena, Epoch AI)
Data sources
- Hugging Face
- GGUF file lists and sizes, model settings (config.json and GGUF headers), downloads and license tags, from the public API.
- LMArena
- Scores from human head-to-head votes, shown as published in the dataset on Hugging Face (CC BY 4.0).
- Epoch AI
- The Epoch Capabilities Index and GPQA Diamond accuracy, shown as published (CC BY 4.0).
- llama.cpp
- The quantization quality table (KL divergence, MIT license), and the reference we check the memory calculation against.
- Hardware specs
- Memory sizes and bandwidth of GPUs and Macs come from the manufacturers' published specs. Each device page shows the source and the date we checked it.
How often it updates
- Every day at 4:20 (Japan time) we refresh from Hugging Face and rebuild the memory estimates and the listing.
- Third-party evaluations (LMArena, Epoch AI) are imported once a week.
- Pages and sitemap.xml show the date each page's content last changed.
How memory is calculated and checked
Required memory is estimated as file size + KV cache + compute buffer + headroom, and the whole formula is published. We checked that the KV cache matches, to the byte, what llama.cpp (b11639) actually allocates, for every architecture we list. The compute buffer and headroom remain estimates.
Editorial policy
- Measured values (file sizes) and estimates (memory, speed) are always labelled differently on the page.
- We do not make our own performance scores or claims like "the best model". Performance figures are third-party values, shown with their source.
- We list models that are used above a threshold and whose memory we can calculate. Modified versions (abliterated, uncensored, etc.) are listed and labelled as such.
- We do not redistribute or mirror model files. Downloads link to the distributor on Hugging Face.
- License summaries are written from the original text, but they are not legal advice.
Who runs this site
Operated by Libra
localllm (localllm.az-arte.com) is run by az-arte.com. The site uses advertising (Google AdSense) and analytics (Google Analytics). To report an error or request a model, please use the contact form.