FinGPT Review in 2026
9 min read

FinGPT is an MIT licensed open source project from the AI4Finance Foundation that fine-tunes open models, mainly Llama 2, for financial sentiment, entity extraction and a weekly stock movement forecaster. It is a strong, cheap research and prototyping base with about 21,600 GitHub stars, but it is not a production document or deal workflow tool, and its adapters inherit the Llama 2 license.
What this covers
- FinGPT is a set of LoRA adapters, datasets and notebooks, not a single model.
- The code is MIT licensed; the Llama 2 base models carry Meta's Community License and Acceptable Use Policy.
- FinGPT v3.3 reports 0.882 weighted F1 on FPB sentiment after a 17.25 hour run on one RTX 3090.
- FinGPT-Forecaster has no published accuracy yet; a historical backtest was merged on 18 November 2026 with live scores pending.
- The FAQ states that extracting structured data from financial reports is not its primary function.
- Best fit: quant and data science research. Poor fit: CIM review, data rooms and IC memos.
FinGPT is an open source research project from the AI4Finance Foundation that fine-tunes existing open models, mainly Llama 2, for financial text tasks like news sentiment, headline classification, entity extraction and a weekly stock movement forecaster. In 2026 it is a well documented, MIT licensed toolkit with about 21,600 GitHub stars, but it is a research and prototyping base, not a production system a fund can switch on.
This review covers what the FinGPT GitHub repository actually contains, what the licenses allow, what it costs to run, and where it fits for a fund, family office or deal team. Every fact below comes from the project's own repository, its Hugging Face pages, its papers, or the Llama 2 license, all fetched on 19 November 2026.
What FinGPT is
FinGPT describes itself as an open source financial large language model project developed and maintained by the AI4Finance Foundation. The repository was created on 11 February 2023. On the day we checked it showed 21.6k stars, 3.1k forks, 36 contributors and 737 commits, with the most recent merge on 18 November 2026. There is one tagged release, FinGPT v1.0.0, dated 8 April 2026.
The important thing to understand is that FinGPT is not one model. It is a collection of LoRA adapters, datasets, notebooks and pipelines. LoRA (low-rank adaptation) is a cheap fine-tuning method that trains a small set of extra weights on top of a base model instead of retraining the whole thing. The project's founding argument, set out in its README and in the first FinGPT paper on arXiv, is that finance data changes too fast to justify BloombergGPT-style training runs, so lightweight adaptation of open models is the practical route.
The README frames the stack as five layers: data source, data engineering, LLMs, task and application. In practice the parts you can download and use today are the task-specific adapters and the datasets that trained them.

AI for finance teams
Screen more deals without adding analysts
First system live in about six weeks. You own the code.
Book a discovery callWhat it actually does
The repository and the FinGPT Hugging Face organisation contain four main pieces of working material.
Sentiment models. The FinGPT v3 series are LoRA fine-tunes on news and tweet sentiment data. v3.3 uses Llama 2 13B as the base, v3.2 uses Llama 2 7B and v3.1 uses ChatGLM2 6B. This is the most mature part of the project and the one with the clearest published benchmarks.
Multi-task adapters. The project released LoRA adapters on Llama 2 7B, Falcon 7B, Bloom 7B1, MPT 7B, ChatGLM2 6B and Qwen 7B, plus a Llama 3 8B multi-task adapter listed on Hugging Face. They cover sentiment, relation extraction, headline classification and named entity recognition.
FinGPT-Forecaster. A Llama 2 7B chat model fine-tuned with LoRA on DOW 30 data. You give it a ticker, a date, a number of past weeks of news and optionally basic financials. It returns positive developments, potential concerns, and a prediction of next week's price movement with a short analysis. The Forecaster README now also lists Indian NSE and BSE symbols as inputs.
Datasets and frameworks. Instruction datasets on Hugging Face include a 76.8K row sentiment training set, 82.2k rows of headline analysis and 17.1k rows of financial Q&A. There is also a FinGPT-RAG framework for retrieval-augmented sentiment analysis, a FinNLP data pipeline and a benchmark suite.
How well it performs
The published numbers are for sentiment classification, and they are respectable for the cost. The table below is taken from the FinGPT README. Scores are weighted F1 on four public financial sentiment datasets.
| Model | FPB | FiQA-SA | TFNS | NWGI | Training hardware and time |
|---|---|---|---|---|---|
| FinGPT v3.3 (Llama 2 13B) | 0.882 | 0.874 | 0.903 | 0.643 | 1 x RTX 3090, 17.25 hours |
| FinGPT v3.2 (Llama 2 7B) | 0.850 | 0.860 | 0.894 | 0.636 | 1 x A100, 5.5 hours |
| OpenAI fine-tune | 0.878 | 0.887 | 0.883 | not reported | not reported |
| GPT-4 | 0.833 | 0.630 | 0.808 | not reported | not reported |
| BloombergGPT | 0.511 | 0.751 | not reported | not reported | 512 x A100, 53 days |
Two caveats matter. First, these are the project's own reported results, and the comparison models are from 2023. GPT-4 is no longer the frontier, so the table tells you FinGPT is competitive on narrow sentiment labelling, not that it beats current general models. Second, sentiment labelling on headlines is a long way from the work a deal team or investment committee actually does.
For the Forecaster there is no published accuracy figure. The README says plainly that results inferred from randomly chosen news can be strongly biased. A pull request merged on 18 November 2026 adds a historical backtest command that reports directional accuracy, mean squared error and parse coverage using only prices available at each forecast cutoff. Its own notes say live scores are still pending. The README also says no validated reinforcement learning trainer for stock prices is included. Until that backtest produces numbers, treat the Forecaster as a demo of the format, not evidence of edge.
The license, and what it allows commercially
The FinGPT code is under the MIT License, copyright 2024 AI4Finance Foundation Inc. MIT allows commercial use, modification and redistribution, provided the copyright and permission notice are kept, and it comes with no warranty.
That is not the whole picture, for two reasons.
First, the README adds a disclaimer that the code is shared "for academic purposes under the MIT education license" and that nothing in it is financial advice or a recommendation to trade real money. The LICENSE file itself is standard MIT and contains no academic restriction, so legally the file governs. The disclaimer still tells you how the maintainers see the project: research, not a supported product.
Second, and more important, the adapters only work on top of a base model, and the base model carries its own license. The setup guide states that the sentiment adapter is a LoRA adapter, not a complete model, and needs the matching Llama 2 13B base. The Forecaster model card lists the adapter as Apache 2.0 but its base as Meta's Llama 2 7B chat model, which is gated on Hugging Face. The Llama 2 Community License grants a royalty-free right to use and modify the model, and commercial use is allowed, with conditions. You must follow Meta's Acceptable Use Policy. You may not use Llama 2 outputs to improve any other large language model. Companies with more than 700 million monthly active users need a separate license from Meta. For a fund or family office the 700 million clause will not bite, but the acceptable use and output restrictions should go past counsel before anything touches client work.
If you use an adapter built on ChatGLM2, Qwen or another base, you inherit that base model's terms instead. Check each one separately.

Build library
See what we have already shipped
Working systems, the stack behind them, and what they cost to run.
Open the libraryWhat it costs to run
The software is free. The costs are compute, data and, optionally, a model API.
Compute. The project's FAQ and setup guide list inference with a cloud API on a CPU with 8GB of RAM, fine-tuning small models on an NVIDIA GPU with 12GB or more of VRAM, and 24GB or more of VRAM for 13B models, with 32GB of system RAM and 50GB or more of SSD storage. The README prices the v3.3 sentiment run at $17.25 on a single RTX 3090, and estimates A100 time at about $4.10 per GPU hour based on the AWS p4d.24xlarge rate of $32.773 per hour as of July 2023. Those are 2023 prices, so reprice them against whatever cloud you use today. Supported Python versions are 3.8 to 3.11, and the guide warns of limited Windows and macOS support for some quantisation libraries.
Data. The Forecaster pulls company profiles, news and financials through yfinance and Finnhub, and needs a Finnhub API key. Building your own Forecaster training set also needs an OpenAI key, because the data preparation notebook uses GPT-4 to generate training labels. An optional sentiment overlay from a provider called Adanos needs its own key. FinGPT does not publish what those third-party services cost; check each provider's pricing for the tier you would need.
Model API. If you skip local models, FinGPT can route inference to a cloud provider. The README lists OpenAI GPT-3.5-turbo as the default fallback and MiniMax as another option. At that point you are paying for, and sending data to, a third-party model.
In short, a working sentiment experiment costs little. A production system costs what any production system costs: engineering time, monitoring, data licensing and someone accountable for the outputs.
Who should use FinGPT, and who should not
FinGPT is a good fit for quant researchers and data scientists who want a clean, reproducible starting point for financial text classification. It is useful for teams who want to learn LoRA fine-tuning on finance data with working notebooks, and for anyone who needs labelled financial instruction datasets to fine-tune a newer base model. The FAQ explicitly encourages using the datasets with other models such as Mistral or Llama 3.
It is a poor fit for anyone looking for a document workflow. The FAQ says directly that extracting structured financial information from reports is not FinGPT's primary function. It does not read a CIM, reconcile a data room, spread financials into a model or draft an IC memo. The fine-tuned RAG models are not released either, only the framework. Most adapters sit on 2023-era 7B and 13B base models, and the Llama 2 base used by the sentiment and Forecaster models has a 4,096 token context window, which is short for long financial documents. Live trading is explicitly out of scope according to the maintainers' own disclaimer.
Where it fits for a fund, family office or deal team
For most investment teams the honest answer is that FinGPT is a component library, not a tool. The sentiment adapters can be one signal inside a news monitoring pipeline for a portfolio or watchlist. The instruction datasets can help evaluate whatever model you do deploy. The Forecaster is worth reading for its prompt structure, which turns news and financials into positives, concerns and a call, a pattern that carries over to screening memos.
The work that eats deal team hours sits elsewhere: reading CIMs and data rooms, checking that numbers tie across documents, building comps, and drafting the memo that goes to committee. That is the territory covered on our AI financial due diligence and AI for private equity pages. Those workflows need long-context models, access to your own documents and permissions, an audit trail for every number, and someone who owns the result when it is wrong. None of that ships in the FinGPT repository.

Implementation services
Put AI into the workflow you actually run
Diligence, close, reconciliation, reporting — in your environment.
See how it worksWhen an in-environment build is the better answer
If the goal is research on financial NLP, clone FinGPT and start with the beginner notebooks. If the goal is a system your team relies on every week, the question changes. You need it to run inside your own environment, on your documents, with current models, and you need to know who maintains it when a data provider changes an endpoint.
That is the work we do. consultance.ai builds custom finance AI systems inside the client's environment, with the first system live in about six weeks, and the client owns the code. Open source projects like FinGPT are often useful parts of that build, as a dataset, a benchmark or a classifier. They are rarely the whole answer. If you are weighing a vendor seat against an owned build, our AI deal desk systems page shows what that looks like, or you can book a call to talk through a specific workflow.
Sources
- AI4Finance-Foundation/FinGPT GitHub repository and README, fetched 19 November 2026
- GitHub API repository record (stars, forks, license, dates), fetched 19 November 2026
- FinGPT LICENSE file, fetched 19 November 2026
- FinGPT FAQ, fetched 19 November 2026
- FinGPT setup guide, fetched 19 November 2026
- FinGPT-Forecaster README, fetched 19 November 2026
- Pull request 283: historical FinGPT Forecaster backtest, fetched 19 November 2026
- FinGPT-Forecaster model card on Hugging Face, fetched 19 November 2026
- FinGPT organisation on Hugging Face (model list), fetched 19 November 2026
- Yang, Liu and Wang, FinGPT: Open-Source Financial Large Language Models, arXiv:2306.06031, fetched 19 November 2026
- Llama 2 Community License Agreement, fetched 19 November 2026
