The three questions everyone asks before running Llama locally.
Which quantization fits my GPU, how much quality does it cost me, and what will inference actually run per month? Llama Local Lab answers all three from published architecture specs — then lets you load a real model into this tab and check the answer yourself.
No account. No API key. No request ever leaves your machine.
Will it fit?
Size every GGUF quant against your VRAM, with the KV cache and compute buffers counted — not just the weights.
What does it cost?
Throughput derived from memory bandwidth, priced against RunPod, Vast.ai, Lambda and owning the hardware.
How do I ship it?
Exact convert, quantize and serve commands for llama.cpp and Ollama. Then evaluate the result in this tab.
01 · Catalog
Llama model catalog
Every Llama checkpoint the studio can plan for, with the architecture numbers the rest of the tools depend on. Pick one and it stays selected across the quantization planner, the command generator and the cost model.
Fetching Hub stats
Family
Showing 13 of 13 checkpoints
Active model meta-llama/Llama-3.1-8B-Instruct
Llama 3.1Text8 languages
Parameters
405.9B
Context
128K tokens
Released
Jul 23, 2024
License
Llama 3.1 Community License
The largest dense open-weights Llama. Frontier-class quality, but needs a multi-node deployment or aggressive quantization to serve.
4% of weights run per token Memory must hold every expert; decode speed tracks the active slice.
The flagship Llama 4 release. 128 experts with 17B active parameters, tuned for high-quality chat, reasoning and image understanding. Targets a single H100 DGX host.
16% of weights run per token Memory must hold every expert; decode speed tracks the active slice.
Natively multimodal mixture-of-experts model. 16 experts, 17B active parameters, and a 10M-token context window — the longest of any open-weights model. Designed to fit on a single H100 at Int4.
Instruction-tuned text model that matches Llama 3.1 405B quality on most benchmarks at a fraction of the serving cost. Still the default choice for self-hosted dense inference.
Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.
Architecture figures (layers, hidden size, head counts, vocabulary) come from each repository’s config.json and Meta’s model cards, bundled with this app so the catalog works with no network at all. Download and like counts are fetched from the public HuggingFace API in your browser — no API key, no proxy — and cached in localStorage for six hours; HuggingFace counts downloads over a rolling 30-day window. Context lengths are the maximum the checkpoint supports, not what your hardware can serve; the quantization planner sizes the KV cache for the window you actually intend to run.