facebook/opt-125m · Hugging Face

Overview

facebook/opt-125m is a lightweight, open-source causal language model developed by Meta AI and hosted on the Hugging Face Hub. With approximately 125 million parameters, it belongs to the OPT (Open Pre-trained Transformer) family introduced in the paper "OPT: Open Pre-trained Transformer Language Models" (arxiv:2205.01068). The model is designed for English text generation tasks and is built on the decoder-only Transformer architecture, making it suitable for a range of natural language processing applications including completion, continuation, and few-shot prompting.

The model was first uploaded to Hugging Face on May 11, 2022, and was last modified on September 15, 2023. It has accumulated over 201 million total downloads across all time, with approximately 17 million downloads recorded in the most recent collection period, reflecting its widespread adoption as an entry-point model for experimentation and education in large language model research.

Capabilities

facebook/opt-125m is configured as an AutoModelForCausalLM with an AutoTokenizer, supporting the standard text-generation pipeline in the Hugging Face Transformers ecosystem. It operates with the following token configuration: the beginning-of-sequence (BOS) and end-of-sequence (EOS) tokens are both set to </s>, while the padding token is <pad>. Token normalization is enabled by default, with no leading or trailing whitespace stripping.

The model supports inference across three major deep learning frameworks: PyTorch, TensorFlow, and JAX, making it accessible to a broad range of developers and researchers. Quantized versions of the model are also available on the Hub, enabling deployment on resource-constrained hardware. The model is tagged for deployment on Microsoft Azure and is region-hosted in the United States.

As a small-scale language model, facebook/opt-125m is particularly well-suited for few-shot learning scenarios, as its architecture and training methodology are rooted in the principles outlined in "Language Models are Few-Shot Learners" (arxiv:2005.14165). It can generate coherent English text continuations from prompt inputs and is commonly used for benchmarking, prototyping, and educational purposes.

Use cases

• Educational and research prototyping: Its small size makes it ideal for students and researchers learning about LLM architecture, fine-tuning, and inference without requiring expensive GPU infrastructure.
• Few-shot prompting experiments: The model supports in-context learning patterns, allowing users to demonstrate few-shot capabilities with minimal computational overhead.
• Text completion and generation: Suitable for simple text continuation tasks, creative writing assistance, and baseline generation benchmarks.
• Pipeline testing: Developers can use it to validate inference pipelines, tokenization workflows, and deployment configurations before scaling to larger models.
• Edge and embedded deployment: Quantized variants enable deployment on devices with limited compute and memory resources.

License & deployment

The model is released under an "other" license as designated by Hugging Face. This typically corresponds to Meta's custom open-research license for the OPT family, which permits research and commercial use under specific conditions. Users should review the LICENSE.md file on the model card for exact terms.

Deployment options include:

• Hugging Face Inference Endpoints: Deploy the model on dedicated, fully managed infrastructure.
• Text Generation Inference (TGI): Serve the model using Hugging Face's optimized inference toolkit.
• Microsoft Azure: The model is tagged for Azure deployment, enabling cloud-based inference.
• Local inference: Run the model locally using the Transformers library with PyTorch, TensorFlow, or JAX backends.

Hardware requirements are modest given the 125M parameter count. Inference can be performed on CPU for basic use cases, while GPU acceleration is recommended for latency-sensitive applications. Quantized versions (e.g., 4-bit or 8-bit) are available for further resource optimization.

Alternatives

• facebook/opt-6.7b / opt-1.3b: Larger variants in the same OPT family offering improved capability at the cost of increased compute requirements.
• facebook/opt-imx-max: Instruction-tuned OPT models designed for chat and conversational applications.
• mistralai/Mistral-7B-v0.1: A more capable open-weight model with strong multilingual and reasoning performance.
• meta-llama/Llama-2-7b-hf: Meta's subsequent open-weight LLM family with improved training data and licensing.
• gpt2: OpenAI's earlier 124M-parameter model, often used as a baseline for small-scale text generation.

FAQ

What is facebook/opt-125m? It is a 125-million-parameter causal language model from Meta AI, part of the OPT open-source model family, designed for English text generation.

What license does it use? The model uses Meta's custom open-research license, listed as "other" on Hugging Face. Consult the LICENSE.md file for full terms.

Can I use it commercially? Commercial use is permitted under Meta's OPT license terms, but users should verify the specific license file for any restrictions.

What frameworks are supported? PyTorch, TensorFlow, and JAX are all supported through the Hugging Face Transformers library.

Is the model quantized? Quantized versions are available on the Hugging Face Hub, enabling deployment on hardware with limited memory.

How does it compare to GPT-2? Both are approximately 125M parameters, but OPT uses a larger training dataset and a slightly different architecture. OPT is generally considered to have stronger few-shot capabilities.

Where can I deploy it? You can deploy it via Hugging Face Inference Endpoints, Text Generation Inference, Microsoft Azure, or run it locally on your own hardware.