Llama models: every size, compared (1B to 405B and Llama 4)

Meta's Llama family runs from a 1.24B model that fits a phone to a 405.85B model that wants a server rack, and the size most people should pick is the 8B class: a 16.07 GB download that runs on an 8 GB graphics card once quantized. Llama 4 Scout and Maverick are the mixture-of-experts outliers, with 17B parameters active per token.

One fact does most of the deciding, and most comparison pages never state it. Released Llama weights are BF16, 2 bytes per parameter, so the download size is also the memory the model needs unquantized: an 8B model is a 16.07 GB download and roughly a 16 GB footprint, and the same weights at Q4 run in about 5 GB. The table below puts every size side by side with the numbers that decide; what this site carries is in the models catalog.

The Llama family in 2026, one table

The ladder, per generation: Llama 2 in 7B, 13B, and 70B; Llama 3 in 8B and 70B; Llama 3.1 in 8B, 70B, and 405B; Llama 3.2 in 1B and 3B text plus 11B and 90B vision; Llama 3.3 in 70B; Llama 4 in Scout and Maverick. The meta-llama README is the source of record for the ladder, the license files, and the official download flow.

Params is the exact count behind the rounded name: 8B is really 8.03B. Files counts the safetensors release set (manifest list for the archive rows). Download size is safetensors bytes. VRAM class applies the help page's quant rules and names the class that runs each model at Q4, with the Q8 class or the qualifying detail in parentheses. License is the label as recorded upstream (the Llama 4 cards record other while naming the Llama 4 Community License Agreement; the cell shows the agreement); what each label requires is the licenses guide's job. The archive column links model pages, or the request queue. Byte totals: Hugging Face listings and the signed manifest. The live catalog is the models list.

What each size is for, honestly

1B and 3B, Llama 3.2. Edge sizes. The 1B is a 2.48 GB download that runs on a phone-class chip or a plain CPU; the 3B at 6.43 GB still fits any laptop. Pick them when latency or battery matters more than peak quality.

8B, Llama 3.1. The workhorse, and the default answer for one gaming GPU: about 5 GB at Q4, so an 8 GB card runs it with room for context, and Q8 wants 12 GB. Meta called the 405B "the first frontier-level open source AI model" in the 3.1 announcement, which upgraded the 8B and 70B and set a 128K context window.

11B and 90B Vision, Llama 3.2. Image and text in, text out, same memory math as the text sizes plus a vision encoder: 21.35 GB and 177.20 GB downloads.

70B, Llama 3.1 and 3.3. The quality step. Q4 is about 43 GB, so the class that runs it is 48 GB: two 24 GB cards, one 48 GB workstation card, or a 48 GB unified-memory Mac. Llama 3.3 exists because Meta claims similar performance compared to the Llama 3.1 405B model at 70B cost; a vendor claim, but why 3.3 replaced 3.1-70B.

405B, Llama 3.1. Server class: 811.72 GB in safetensors, about 243 GB at Q4, reference deployment 8x A100 80 GB. The Llama 3 paper is the record for the model itself, a 405B dense model with a 128K context window.

Scout and Maverick, Llama 4. Mixture-of-experts: 109B total parameters for Scout, about 400B for Maverick, and only 17B active per token, so generation runs at small-model speed while memory stays large-model. Scout's model card says it fits a single H100 (80 GB) with int4 quantization and claims a 10M token context; Maverick claims 1M. The larger Behemoth was announced and never released.

Download size is not VRAM (but they are related)

The download column and the VRAM column measure the same weights in two states. Released Llama weights are BF16 at 2 bytes per parameter, so loaded unquantized, bytes on disk equal bytes in memory: 16.07 GB downloaded is about 16 GB held. Quantization splits the numbers; a Q4 copy needs roughly 0.6 GB per billion parameters at runtime, so the same 8B model runs in about 5 GB.

You still download the format you chose. A Q4 GGUF of the 8B is a 4.9 GB download, the 70B is 43 GB, and the 405B is 243 GB, per ollama's llama3.1 and llama4 library pages. Those are repack bytes: someone converted the safetensors release to GGUF and quantized it, and you download their conversion. This archive mirrors the upstream safetensors verbatim; if a GGUF is what you want, the verify guide shows how to check one.

Context costs memory on top of weights: plan 1.2x to 1.5x the model size for weights plus working context, a community rule of thumb. And the biggest repo hides a trap unrelated to quantization.

Which size fits your hardware

Apply the help page's rules and the classes fall out: roughly 0.6 GB per billion parameters at Q4, 1 GB per billion at Q8, and when a card sits between sizes, start one smaller than you think fits.

Llama, llama.cpp, ollama: three different things

Search results mix three unrelated things: the llama.cpp home page outranks model pages for a sizes query on some engines, and "download llama.cpp" tops the autocomplete for "download llama". The split:

• Llama is Meta's model family: the thing with sizes, licenses, and weight files.
• llama.cpp is an open-source inference engine, a program that loads model files and runs them on your own hardware. Downloading it gets you software, not a model.
• ollama is a client built on that stack which fetches quantized copies for you; its library pages are the source of blob sizes like the 4.9 GB 8B. Client behavior, including why a pull crawls, is the ollama guide's territory.

Third-party fine-tunes such as Hermes 3 and saiga are Llama-architecture models: another lab trained on Meta's released weights, so they are Llama by architecture, not Meta releases, and the base license stays attached.

Where to download each one

Four routes cover the family. They differ less in bytes than in gating.

The gate is the real difference. Every repository under Meta's own org requires access approval after you accept terms; the third-party fine-tunes do not. Hermes 3 and saiga_llama3_8b are open downloads on Hugging Face, and both are mirrored here. Meta's own weights are not carried today.

What the archive carries today, and how to check it

The Llama-family models in the archive today:

• NousResearch_Hermes-3-Llama-3.1-8B, a Llama 3.1 8B fine-tune from Nous Research. 11 files, 16,069,735,126 bytes, license label llama3, revision 896ea440e5a9e6070e3d8a2774daf2b481ab425b pinned in the signed manifest. Upstream: the Hermes 3 model page.
• IlyaGusev_saiga_llama3_8b, a Russian-language fine-tune of Llama 3 8B Instruct. 11 files, 16,069,730,325 bytes, revision 5bb9917bdb85340549662ebb62c8e522037ff3f3. Upstream records the license as other, and this site records it the same way; the base model carries the Llama 3 Community License, and the licenses guide covers what that obligates. Upstream: the saiga model page.

Both payloads have the shape of Meta's own 8B release: four safetensors shards, three near 4.9 GB and one near 1.1 GB, plus tokenizer and configs. Both were verified against upstream metadata at fetch time on 2026-08-23.

Every payload in this archive was verified against upstream at fetch time, with the upstream revision pinned in the signed manifest, and the per-file digests ride along: sha256 for LFS files, sha1+size for git blobs, each labeled with its method. The verify page has the command card, the verify-model-download guide walks every file of a real model, and the magnet links and infohashes guide explains the vocabulary. The AI model torrents guide tells the distribution story once.

Pick by hardware: the largest size your card holds at Q4, then Q8 if you have room. Download the format you intend to keep, verify it, and leave the client running so the next person pulls from a swarm instead of one server; the contribute page covers the rest.