FinGPT

What it is

FinGPT is the AI4Finance Foundation's open financial LLM project, and what you download is weights and data rather than an application: eleven LoRA adapters on Hugging Face — fine-tunes of Llama-2, Llama-3, ChatGLM2, Falcon, MPT, Bloom, Qwen and InternLM for sentiment, named-entity recognition, relation extraction and headline classification — fourteen instruction datasets, and the notebooks that trained and benchmarked them.

The repository is alive; the modelling is not. MIT per its LICENSE file, 21.2k stars, not archived, last commit to master on 12 September 2026, one tagged release (v1.0.0, 8 April 2026) — but the newest adapter was last touched in October 2024 and the rest are from 2023 and the first half of 2024. Over the past 30 days the instruction datasets pulled thousands of downloads, 2,015 for fingpt-sentiment-train alone, while eight of the eleven adapters pulled none. People take the data and train their own.

Pricing

The repository and the weights are MIT, commercial use included. The project site advertises "API & Service Plans" on top — Free at $0 capped at 100 sentiment requests a day, Premium at $19.90 a month for 10,000, Enterprise on request — with nothing behind them: no API host resolves, the documentation page's "API Reference" leads back to the README, the advertised demo path returns 404, and "Start Premium" is a jump link to a contact block whose only action is an email address. Budget for GPU time, not for a plan.

Data & coverage

FinGPT ships no market data. Its datasets are labelled financial text — sentiment over news and tweets, headline classification, NER, relation extraction, FiQA Q&A and a Chinese multiple-choice set — covering US news and Chinese A-share material. The most application-shaped piece, FinGPT-Forecaster, is a Llama-2-7b LoRA trained on Dow 30 news to 2023 that emits a next-week price-movement call from Finnhub and yfinance data you fetch with your own keys — its own README calls it a "junior robo-advisor", which is why this card sets ai_features: recommendations.

Integrations

Python, and a 2023 Python: the Forecaster pins torch 2.0.1, transformers 4.32.0, peft 0.5.0 and gradio 3.50.2, and the setup guide warns off 3.12. Llama-2 base weights are gated, so you also need a Hugging Face token. The README advertises a no-GPU path — set FINGPT_LLM_PROVIDER=openai and import fingpt.Forecaster — but no such module exists in the repository, and the only code that reads that variable belongs to an unrelated sub-project committed alongside FinGPT.

Limitations

• The published weights are two years old, on base models that have been superseded. What is still worth having is the datasets and the training recipe, not the adapters.
• The hosted demo does not run: the Forecaster Space sits in a runtime error — importing peft pulls in bitsandbytes, which cannot import driver from triton.runtime — and the Chinese-language Space is paused.
• No packaging worth the name — the PyPI entry is a stub at 0.0.1 from October 2023, with no console entry point — so you vendor the repository or copy the notebook.
• Fine-tuning needs an NVIDIA card with 12GB+ of VRAM; the headline v3.3 sentiment model took 17.25 hours on one RTX 3090.
• Nothing here is inference infrastructure: no server, no API, no evaluation harness you would put in front of production traffic.

Alternatives

FinBERT is smaller and simpler if sentiment is all you want. FinRobot, from the same foundation, is the thing that runs rather than the thing you fine-tune. And a current general-purpose model with a decent prompt will beat these 2023 adapters on most of the tasks they were built for.