LL

Meta-Llama-3-120B-Instruct

LlamaGGUF by bartowskiUpdated 2024-05-09
Download on Hugging Face
Parameters122B
Context8K
Layers140
Architecturellama
Smallest quant26.0 GB
Quick answer

Meta-Llama-3-120B-Instruct needs about 78.4 GB of memory at Q4_K_M (a 73.2 GB file) with an 8K context window, which is more than a 48 GB GPU holds. This repository has 23 quants, from 26.0 GB (IQ1_S) to 243.8 GB.

Meta‑Llama‑3‑120B‑Instruct – 120 B parameter instruction model

Meta‑Llama‑3‑120B‑Instruct is a 120 B parameter Llama architecture model trained on 8192 context tokens. It is available in multiple GGUF quantizations from IQ1_S to Q8_0 and an unknown quant. The model is released under an other license.

Good at

  • Large‑scale instruction following
  • High‑capacity context handling
  • Multilingual text generation

Limits

  • No vision support
  • Requires large VRAM for higher quants
  • License not specified for commercial use

Best for

  • Enterprise chatbots
  • Research prototypes
  • Long‑form content creation

Who it’s for. Developers and researchers needing a high‑parameter instruction model for text generation. The model is suitable for environments with sufficient GPU memory and where the license terms are acceptable.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ1_S26.0 GB 31.2 GB
Too big Get file →
IQ1_M28.5 GB 33.7 GB
Too big Get file →
IQ2_XXS32.6 GB 37.8 GB
Too big Get file →
IQ2_XS36.2 GB 41.4 GB
Too big Get file →
IQ2_S38.0 GB 43.2 GB
Too big Get file →
IQ2_M41.3 GB 46.5 GB
Too big Get file →
Q2_K45.1 GB 50.3 GB
Too big Get file →
IQ3_XXS47.0 GB 52.2 GB
Too big Get file →
IQ3_XS50.2 GB (2 parts) 55.3 GB
Too big Get file →
Q3_K_S52.8 GB (2 parts) 58.0 GB
Too big Get file →
IQ3_S53.0 GB (2 parts) 58.2 GB
Too big Get file →
IQ3_M54.7 GB (2 parts) 59.9 GB
Too big Get file →
Q3_K_M58.8 GB (2 parts) 64.0 GB
Too big Get file →
Q3_K_L64.0 GB (2 parts) 69.2 GB
Too big Get file →
IQ4_XS65.3 GB (3 parts) 70.5 GB
Too big Get file →
IQ4_NL69.0 GB (3 parts) 74.2 GB
Too big Get file →
Q4_K_S69.4 GB (3 parts) 74.6 GB
Too big Get file →
Q4_K_M73.2 GB (3 parts) 78.4 GB
Too big Get file →
Q5_K_S84.0 GB (3 parts) 89.2 GB
Too big Get file →
Q5_K_M86.2 GB (3 parts) 91.4 GB
Too big Get file →
Q6_K100.0 GB (4 parts) 105.2 GB
Too big Get file →
Q8_0129.5 GB (5 parts) 134.7 GB
Too big Get file →
—243.8 GB (8 parts) 249.0 GB
Too big Get file →

For 8 GB VRAM, use IQ1_S (26.03 GB file, 31.2 GB memory at 8k context). For 12 GB VRAM, IQ1_M (28.5 GB file, 33.7 GB memory). For 16 GB VRAM, Q2_K (45.1 GB file, 50.3 GB memory).

What GPU do you need for Meta-Llama-3-120B-Instruct?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: no quant fits at an 8K context.
  • 16 GB VRAM: no quant fits at an 8K context.
  • 24 GB VRAM: no quant fits at an 8K context.
  • 32 GB VRAM: the best quant that fits is IQ1_S (26.0 GB file, about 31.2 GB in use).
  • 48 GB VRAM: the best quant that fits is IQ2_M (41.3 GB file, about 46.5 GB in use).

Which Meta-Llama-3-120B-Instruct quant should you choose?

Q4_K_M is the usual starting point for Meta-Llama-3-120B-Instruct: it balances file size and quality well. Here it is a 73.2 GB file.

The Q8_0 file is 129.5 GB, about 1.8 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is IQ1_S at 26.0 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters122B
Layers140
Trained context8,192 tokens
KV cache (fp16)560 KB per token

Source

Quantized bybartowski
FamilyLlama
Licenseother
Downloads (HF)639
Last updated2024-05-09

Frequently asked questions about Meta-Llama-3-120B-Instruct

How much VRAM does Meta-Llama-3-120B-Instruct need?

About 78.4 GB at Q4_K_M with an 8K context window. That is the 73.2 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Meta-Llama-3-120B-Instruct on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 31.2 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Meta-Llama-3-120B-Instruct GGUF?

It depends on the quant: IQ1_S is 26.0 GB, IQ1_M is 28.5 GB, IQ2_XXS is 32.6 GB, IQ2_XS is 36.2 GB, IQ2_S is 38.0 GB, IQ2_M is 41.3 GB and 17 more.

What context length does Meta-Llama-3-120B-Instruct support?

The GGUF header lists a trained context length of 8,192 tokens (8K). The KV cache grows with the context you use: about 4.7 GB at 8K and 4.7 GB at 8K.

Where can I download Meta-Llama-3-120B-Instruct GGUF?

The files are on Hugging Face in the bartowski/Meta-Llama-3-120B-Instruct-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Meta-Llama-3-120B-Instruct use?

Hugging Face lists the license as other. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run Q4_K_M on 8 GB VRAM?

No, Q4_K_M requires 78.4 GB memory at 8k context, exceeding 8 GB VRAM.

What is the smallest quant available?

IQ1_S is the smallest quant, 26.03 GB file and 31.2 GB memory at 8k context.

Does the model support vision?

No, the model has no vision capabilities.

Other GGUFs of Meta-Llama-3-70B-Instruct

Similar models

Browse all →

More from bartowski

All by bartowski →

More Llama models

All Llama →

Other large models

All large models →