LL

Meta Llama 3.1 70B Instruct

LlamaGGUF by legraphistaUpdated 2024-07-27
Download on Hugging Face
Parameters70.6B
Context128K
Layers80
Architecturellama
Smallest quant15.3 GB
Quick answer

Meta Llama 3.1 70B Instruct needs about 35.1 GB of memory at IQ3_M (a 31.9 GB file) with an 8K context window, so it fits on a 48 GB GPU. This repository has 25 quants, from 15.3 GB (IQ1_S) to 141.1 GB. Longer context needs more: at 32K tokens the KV cache alone adds about 10.7 GB.

Meta Llama 3.1 70B Instruct quantized for llama.cpp

Meta Llama 3.1 70B Instruct is a 70.6B parameter model with a 131072 token context window. It is available in multiple GGUF quantizations from IQ1_S to Q8_0 and BF16. The model is licensed under llama3.1 and supports instruction following.

Good at

  • Large context generation
  • Instruction following
  • High token throughput

Limits

  • No vision support
  • Requires large VRAM for higher quants
  • License is llama3.1

Best for

  • Chatbots needing long context
  • Research prototyping
  • Fine‑tuning on local hardware

Who it’s for. Users who need a large instruction‑tuned model on local hardware with llama.cpp. Suitable for developers and researchers who can manage the memory requirements of the chosen quantization.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ1_S15.3 GB 18.5 GB
Too big Get file →
IQ1_M16.8 GB 19.9 GB
Too big Get file →
IQ2_XXS19.1 GB 22.3 GB
Too big Get file →
IQ2_XS21.1 GB 24.3 GB
Too big Get file →
IQ2_S22.2 GB 25.4 GB
Too big Get file →
IQ2_M24.1 GB 27.3 GB
Too big Get file →
Q2_K_S24.5 GB 27.7 GB
Too big Get file →
Q2_K26.4 GB 29.6 GB
Too big Get file →
IQ3_XXS27.5 GB 30.7 GB
Too big Get file →
IQ3_XS29.3 GB 32.5 GB
Too big Get file →
IQ3_S30.9 GB 34.1 GB
Too big Get file →
Q3_K_S30.9 GB 34.1 GB
Too big Get file →
IQ3_M31.9 GB 35.1 GB
Too big Get file →
Q3_K34.3 GB 37.5 GB
Too big Get file →
Q3_K_L37.1 GB 40.3 GB
Too big Get file →
IQ4_XS37.9 GB 41.1 GB
Too big Get file →
IQ4_NL40.1 GB 43.2 GB
Too big Get file →
Q4_K_S40.3 GB 43.5 GB
Too big Get file →
Q4_K42.5 GB 45.7 GB
Too big Get file →
Q5_K_S48.7 GB (3 parts) 51.8 GB
Too big Get file →
Q5_K49.9 GB (3 parts) 53.1 GB
Too big Get file →
Q6_K57.9 GB (3 parts) 61.1 GB
Too big Get file →
Q8_075.0 GB (4 parts) 78.2 GB
Too big Get file →
BF16141.1 GB (6 parts) 144.3 GB
Too big Get file →
—141.1 GB (6 parts) 144.3 GB
Too big Get file →

For 8 GB VRAM, use IQ1_S or IQ1_M (file size 15.34 GB or 16.75 GB, memory 18.5 GB or 19.9 GB). For 12 GB VRAM, IQ2_XXS or IQ2_XS are suitable. For 16 GB VRAM, IQ2_S or IQ2_M provide a balance between size and memory. Higher quants require more VRAM and may need split files.

What GPU do you need for Meta Llama 3.1 70B 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: the best quant that fits is IQ2_XXS (19.1 GB file, about 22.3 GB in use).
  • 32 GB VRAM: the best quant that fits is IQ3_XXS (27.5 GB file, about 30.7 GB in use).
  • 48 GB VRAM: the best quant that fits is Q4_K (42.5 GB file, about 45.7 GB in use).

Which Meta Llama 3.1 70B Instruct quant should you choose?

IQ3_M (31.9 GB) is the middle option in this repository. A reasonable choice when a 4-bit file will not fit, mostly on larger models. Quality loss is noticeable but often acceptable.

If memory is tight, the smallest file is IQ1_S at 15.3 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
Parameters70.6B
Layers80
Trained context131,072 tokens
KV cache (fp16)320 KB per token

Source

Quantized bylegraphista
FamilyLlama
Licensellama3.1
Downloads (HF)1,759
Last updated2024-07-27

Frequently asked questions about Meta Llama 3.1 70B Instruct

How much VRAM does Meta Llama 3.1 70B Instruct need?

About 35.1 GB at IQ3_M with an 8K context window. That is the 31.9 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 3.1 70B Instruct on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 18.5 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Meta Llama 3.1 70B Instruct GGUF?

It depends on the quant: IQ1_S is 15.3 GB, IQ1_M is 16.8 GB, IQ2_XXS is 19.1 GB, IQ2_XS is 21.1 GB, IQ2_S is 22.2 GB, IQ2_M is 24.1 GB and 19 more.

What context length does Meta Llama 3.1 70B Instruct support?

The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 2.7 GB at 8K and 10.7 GB at 32K.

Where can I download Meta Llama 3.1 70B Instruct GGUF?

The files are on Hugging Face in the legraphista/Meta-Llama-3.1-70B-Instruct-IMat-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 3.1 70B Instruct use?

Hugging Face lists the license as llama3.1. 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 IQ1_S on 8 GB VRAM?

Yes, IQ1_S requires 18.5 GB memory at 8k context, so it fits in 8 GB VRAM with batching.

How many shards does Q5_K need?

Q5_K is split into 3 shards, each part of the 49.95 GB file.

What is the largest quant that fits in 16 GB VRAM?

IQ2_M requires 27.3 GB memory, so it exceeds 16 GB VRAM; use a lower quant like IQ2_S or IQ2_XS.

Other GGUFs of Llama-3.1-70B-Instruct

Similar models

Browse all →

More from legraphista

All by legraphista →

More Llama models

All Llama →

Other large models

All large models →