The Fin AI

The Fin AI is a global academy advancing financial general intelligence through education, open research and incubation. This is where we release our benchmarks, models, leaderboards and demos for financial AI — across English, Chinese, Japanese, Spanish and Greek, and across text, vision and audio. 🌐 thefin.ai · 📄 all our papers

🎯 Our vision

Open-source, trustworthy Financial General Intelligence — pursued together by researchers, engineers and financial experts around the world. AI Readiness in Finance.

Three pillars, one mission:

🏅 Named one of Global Finance's World's Best Financial Innovation Labs 2026.

News 🗞️

• Oct 2026 · All leaderboards redesigned, with live data and a last updated date; the Plutus-8B and FinLLaVA demos are back on ZeroGPU.
• May 2026 · Herculean: agentic financial analysis — trading, hedging, reporting and auditing.
• Mar 2026 · EnterpriseArena: can LLM agents act as CFOs over long horizons?
• Feb 2026 · Ebisu for Japanese financial language, and Conv-FinRe for conversational financial recommendation.
• Nov 2025 · FinCriticalED: fact-level OCR of financial documents.
• Oct 2025 · FinAuditing for taxonomy-structured auditing, and the live Agent Market Arena.

🏆 Leaderboards & demos

• Open FinLLM Leaderboard — MultiFinBen: 21 models on 36 datasets in five languages and three modalities.
• Reasoning Leaderboard — Fino1: numerical reasoning over reports, long documents and XBRL.
• Greek Financial LLM Leaderboard — Plutus: five Greek financial tasks.
• Agent Market Arena — LLM trading agents competing in live markets.
• IJCAI 2024 FinLLM Challenge — final results of the shared task.
• Demos: Plutus-8B (Greek financial chat) · FinLLaVA (questions about charts and tables)

📊 Benchmarks

Multilingual & multimodal

• MultiFinBen — text, OCR and audio tasks in five languages (arXiv:2506.14028)
• ICE-PIXIU Chinese · FLARE-ES Spanish · Plutus Greek · Ebisu Japanese
• FinCriticalED — fact-level OCR of financial documents (arXiv:2511.14998)

Financial NLP, reasoning & risk

• FinBen & PIXIU — holistic English evaluation: extraction, QA, summarization, forecasting, trading (arXiv:2402.12659)
• CALM — credit scoring, fraud, financial distress and claims (arXiv:2310.00566)
• FMD misinformation detection · Conv-FinRe conversational recommendation

Filings, XBRL & auditing

• FinTagging — XBRL tagging of financial reports (arXiv:2505.20650)
• FinAuditing — taxonomy-structured, multi-document auditing (arXiv:2510.08886)

Agents & decision-making

• Herculean (arXiv:2605.14355) · EnterpriseArena (arXiv:2603.23638) · Agent Market Arena (arXiv:2510.11695)

🧠 Models

• Fin-o1 / Fino1 — reasoning-enhanced financial LLMs, 8B and 14B (arXiv:2502.08127)
• Open-FinLLMs — FinLLaMA, FinLLaMA-Instruct and the multimodal FinLLaVA (arXiv:2408.11878)
• Plutus-8B — the first Greek financial LLM (arXiv:2502.18772)
• FinSeer & StockLLM — retrieval-augmented time-series forecasting (arXiv:2502.05878)
• FinMA — the original PIXIU instruction-tuned models (arXiv:2306.05443)

⚡ Quick start

Most datasets ask for a short form (name, affiliation, intended use) that is approved automatically. Accept it once on the dataset page, log in, and load:

from datasets import load_dataset

ds = load_dataset("TheFinAI/en-fpb")  # FinBen · Financial PhraseBank sentiment

Datasets are named {language}-{task} — en-fpb, zh-stocka, es-tsa, gr-qa, jp-icr, multi-… for multilingual — while paper-named benchmarks keep their names (FinCriticalED, Herculean). Old names such as flare-fpb still redirect. Licenses follow the original sources and are stated on every card. To evaluate a model on our tasks, start from FinBen or MultiFinBen.

🤝 Community

Website · GitHub · Discord · Google Groups · YouTube · Google Scholar

Questions about a dataset or model? Open a discussion on its page.