What are the hardware requirements for Opt 6.7B?

Opt 6.7B requires approximately 13.4 GB of VRAM and supports a context window of 2K tokens. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed Opt 6.7B and how large is it?

Opt 6.7B is developed by facebook, a model with 6.7b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

How does Opt 6.7B perform on standard benchmarks?

Opt 6.7B has the following published scores: MMLU 24.57. Compare against reference models on this page or on the LLM Explorer leaderboards.

Opt 6.7B by facebook — VRAM 13.4GB, 2K context

Name: Opt 6.7B
Author: facebook

Opt 6.7B is an open-source language model by facebook. Features: 6.7b LLM, VRAM: 13.4GB, Context: 2K, License: other, LLM Explorer Score: 0.11, Arc: 39.2, HellaSwag: 68.7, MMLU: 24.6, GSM8K: 1.

Arxiv:2005.14165 Arxiv:2205.01068 En Jax Opt Pytorch Region:us Sharded Tf

Model Card on HF 🤗: https://huggingface.co/facebook/opt-6.7b

Opt 6.7B Parameters and Internals

Model Type

text-generation, decoder-only

Use Cases

Areas:

Research

Applications:

Text Generation, Prompt-based Evaluation

Primary Use Cases:

Text generation using CLM objective

Limitations:

High possibility of bias and quality issues like hallucination and lack of diversity

Additional Notes

OPT models aim to enable reproducible and responsible research.

Supported Languages

English (Predominantly supported), Non-English (Small amount in training corpus)

Training Details

Data Sources:

BookCorpus, CC-Stories, The Pile, Pushshift.io Reddit dataset, CCNewsV2

Data Volume:

180B tokens

Methodology:

Causal Language Modeling (CLM)

Context Length:

2048

Training Time:

~33 days

Hardware Used:

992 A100 80GB GPUs

Model Architecture:

Decoder-only, similar to GPT-3

Responsible Ai Considerations

Mitigation Strategies:

Model may have bias due to unfiltered internet data.

Input Output

Input Format:

Sequences of 2048 consecutive tokens, tokenized using GPT2 BPE with a vocabulary of 50272.

Accepted Modalities:

text

Output Format:

Text

Performance Tips:

Use the generate method directly for better performance with large models.

Release Notes

Version:

175B

Date:

2022-05-03

Notes:

Initial release with sizes from 125M to 175B parameters.

LLM Name	Opt 6.7B
Repository 🤗	https://huggingface.co/facebook/opt-6.7b
Model Size	6.7b
Required VRAM	13.4 GB
Updated	2026-07-29
Maintainer	facebook
Model Type	opt
Model Files	10.0 GB: 1-of-2 3.4 GB: 2-of-2
Supported Languages	en
Model Architecture	OPTForCausalLM
License	other
Context Length	2048
Model Max Length	2048
Transformers Version	4.21.0.dev0
Beginning of Sentence Token	</s>
End of Sentence Token	</s>
Unk Token	</s>
Vocabulary Size	50272
Torch Data Type	float16
Activation Function	relu
Errors	replace

Rank the Opt 6.7B capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Quantized Models of the Opt 6.7B

Model	Likes	Downloads	VRAM
Opt 6.7B GPTQ 4bit G128	0	37	4 GB
Opt 6.7B Gptq 4bit	0	90	4 GB

Best Alternatives to Opt 6.7B

Best Alternatives	Context / RAM	Downloads	Likes
...7B Gqa Ub 16 Best For KV Cache	2K / 27.4 GB	190	0
...B Qcqa Ub 16 Best For KV Cache	2K / 27.4 GB	94	0
....7B Qcqa Ub 16 Best For Q Loss	2K / 27.4 GB	60	0
Galactica 6.7B ReFT GSM8k	2K / 13.8 GB	58	0
Galactica 6.7B EssayWriter	2K / 13.4 GB	55	5
Opt Flan Iml 6.7B	2K / 26.6 GB	131	1
...lactica 6.7B Evol Instruct 70K	2K / 27.4 GB	129	19
OPT 6.7B Nerybus Mix	2K / 13.7 GB	125	20
Galactica 6.7B Finetuned	2K / 27.4 GB	89	34
OPI Galactica 6.7B	2K / 13.4 GB	116	6

Note: green Score (e.g. "73.2") means that the model is better than facebook/opt-6.7b.