OPT
Meta's open-source language model with fully released training code and logbook — complete transparency.
Overview
OPT (Open Pre-trained Transformer) is a suite of decoder-only language models released by Meta AI Research in May 2022, ranging from 125M to 175B parameters. OPT was significant as one of the first near-GPT-3-scale models released publicly, enabling research that had previously been impossible without access to proprietary commercial models.
Uniquely, Meta released not just the model weights but the complete training codebase and a detailed training logbook documenting every decision, setback, and modification made during the training process. This level of transparency was unprecedented for a large model and provided the research community with valuable insights into the practical challenges of training very large language models.
OPT-175B matched GPT-3 on many benchmarks while being fully open for research use. The model enabled research into large model behaviour, few-shot learning, and factual knowledge that required full access to model internals. While OPT has been substantially surpassed by more recent models like Llama 3, it remains historically significant as a transparent large model research artefact.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
ASEAN Perspective
OPT in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
OPT was Meta's 2022 effort to give researchers a GPT-3-scale model, releasing weights up to 175B alongside the training code and a candid logbook of divergences, restarts and mid-run fixes — a transparency milestone the field still cites. For practical use in 2026 it's obsolete: the license is research-only and non-commercial, quality sits far behind modern Llama, Qwen and Mistral (and behind small current open models), and Meta superseded it with the Llama line from early 2023. There's no first-party API — it's checkpoints on Hugging Face, nothing more. It suits reproducibility studies and teaching the history of large-model training, not production. Free for research; the low score reflects current capability and value, not its considerable historical importance.
What people say
The logbook is the reason OPT still gets cited. When Meta AI released the Open Pre-trained Transformer suite in May 2022 — decoder-only models from 125M up to 175B parameters — it shipped not just weights but the full training codebase and a running diary of the training itself: loss divergences, hardware failures, mid-run hyperparameter surgery, restarts. Nobody at GPT-3 scale had shown their work like that before, and LLM courses still assign it as the honest account of what training a very large model actually costs.
As a model, OPT-175B roughly matched GPT-3 on standard benchmarks, and for a research community locked out of proprietary systems that access mattered: work on few-shot behaviour, memorisation and model internals that needed full weight access ran on OPT because nothing else was available at that scale.
None of that makes it usable today. The license is research-only and non-commercial, so it was never a production option even when current. Quality-wise it sits several generations behind Llama, Qwen and Mistral — modern sub-10B open models beat it comfortably — and Meta itself superseded it within nine months when the first LLaMA landed in February 2023. There's no API and no service; it's a set of checkpoints on Hugging Face.
The low score here is a statement about 2026 utility, not historical importance. If you're doing reproducibility work or teaching the history of scaling, OPT is a first-rate artefact. For anything else, it's a museum piece — and the museum built its own successor wing three doors down.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- OPT-175B was the first near-GPT-3-scale model released publicly with full training transparency including a day-by-day training logbook.
- The training logbook candidly described hardware failures, training instabilities, and engineering workarounds — a valuable resource for any team attempting to train large models.
- OPT was trained on 992 A100 80GB GPUs for approximately 33 days, with the full training run costing an estimated $2 million in compute.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including OPT's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with OPT unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to OPT directly →
Spotted something out of date? Suggest an update →
OPT in the news
AI & ML
Meta Ships Muse Spark 1.1 — Its First Frontier API With a Price Tag, and a Shift Away From...
Cloud & Infra
Meta Compute: Meta Is Reportedly Exploring a Cloud Business for Excess AI Capacity
Developer Tools
Miasma: How a Hijacked CI Pipeline Shipped Malicious npm Packages With Valid Provenance
Alternatives to OPT
More in LLMs & Chat
