Skip to the studio
Llama Local Lab
Runs entirely in your browser

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.

Loading HuggingFace statistics

Llama 4MultimodalMoE12 languages
Parameters
400B
Context
1M tokens
Released
Apr 5, 2025
Mixture of experts
Experts
128
Active / token
1
Active params
17B

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.

Loading HuggingFace statistics
Community GGUF mirrors

Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.

Llama 4MultimodalMoE12 languages
Parameters
109B
Context
10M tokens
Released
Apr 5, 2025
Mixture of experts
Experts
16
Active / token
1
Active params
17B

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.

Loading HuggingFace statistics
Community GGUF mirrors

Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.

Llama 3.2Multimodal1 languages
Parameters
88.6B
Context
128K tokens
Released
Sep 25, 2024
License
Llama 3.2 Community License

The large vision model of the 3.2 generation. Needs 2× 80 GB at FP16 or a single 80 GB card at Q4.

Loading HuggingFace statistics

Llama 3.3Text8 languages
Parameters
70.6B
Context
128K tokens
Released
Dec 6, 2024

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.

Loading HuggingFace statistics
Community GGUF mirrors

Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.

Llama 3.1Text8 languages
Parameters
70.6B
Context
128K tokens
Released
Jul 23, 2024
License
Llama 3.1 Community License

The 3.1 generation 70B. Superseded by Llama 3.3 70B, which has the same architecture and better instruction tuning.

Loading HuggingFace statistics

Code LlamaText1 languages
Parameters
33.7B
Context
16K tokens
Released
Aug 24, 2023
License
Llama 2 Community License

Legacy code model kept here as a size reference point — it is the classic "does a 34B fit on my 24 GB card" case.

Loading HuggingFace statistics

Llama 3.2Multimodal1 languages
Parameters
10.6B
Context
128K tokens
Released
Sep 25, 2024
License
Llama 3.2 Community License

Llama 3.1 8B plus a cross-attention vision adapter. Document understanding and chart reasoning on a single 24 GB card.

Loading HuggingFace statistics

Llama 3.1Text8 languages
Parameters
8.03B
Context
128K tokens
Released
Jul 23, 2024
License
Llama 3.1 Community License

The workhorse. Fits comfortably on a single consumer GPU at Q4_K_M and runs at conversational speed on a laptop.

Loading HuggingFace statistics
Community GGUF mirrors

Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.

Llama GuardText1 languages
Parameters
8.03B
Context
128K tokens
Released
Jul 23, 2024
License
Llama 3.1 Community License

Safety classifier fine-tuned from Llama 3.1 8B. Run it alongside your generator to label prompts and responses against the MLCommons taxonomy.

Loading HuggingFace statistics

Llama 3.2Text8 languages
Parameters
3.21B
Context
128K tokens
Released
Sep 25, 2024
License
Llama 3.2 Community License

Edge-class model with tied embeddings. Strong summarization and rewriting for its size; runs in a browser tab via WebGPU.

Loading HuggingFace statistics
Community GGUF mirrors

Community conversions, not published by Meta. Check the quant list and the file checksums before you run one.

Llama GuardText1 languages
Parameters
1.50B
Context
128K tokens
Released
Sep 25, 2024
License
Llama 3.2 Community License

Pruned and distilled safety classifier small enough to run on-device next to a 3B generator.

Loading HuggingFace statistics

Llama 3.2Text8 languages
Parameters
1.24B
Context
128K tokens
Released
Sep 25, 2024
License
Llama 3.2 Community License

The smallest official Llama. Roughly 20% of its parameters are the 128k-token embedding table, which is why naive size formulas get it wrong.

Loading HuggingFace statistics
Community GGUF mirrors

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.