Open source · MIT · 100% local

Wall Street's tools.
Your GPU.
Open source.

FNTK Pro Local is an open-source hedge fund analytics terminal. Ask SEC filings, earnings calls, and market data anything in plain English — answered by an LLM running on your machine. No cloud. No accounts. No telemetry.

Run it yourself ↗ pip install git+https://gitlab.com/standardhorizons/fntkprolocal
A home GPU cluster glowing green and magenta in a dark garage, monitor showing candlestick charts, punk stickers on the wall
fig. 1 — the new trading floor (your garage)
// capabilities

What it does out of the box

01

Ask the filings

Ask 10-Ks, S-1s, proxy statements, and 8-Ks anything in plain English. "What was NVDA's data center revenue growth?" — answered from the actual filing.

02

Earnings-call Q&A

Query earnings-call transcripts directly. "What guidance did management give on margins?" Get the exact quote.

03

Find the next one

Local vector search over 10-K product descriptions. "Find companies similar to CELH." No screener subscription needed, and nothing leaves your machine.

04

Write the memo

Investment-memo workflows with side-by-side comparison. Draft the thesis, stress-test it, diff the takes — all on your hardware.

05

Market data on demand

Price moves, earnings calendars, 52-week lows, institutional holders, biotech pipelines. Plain JSON POST endpoints. Pipe them anywhere.

06

Automate the research

Chain multi-step research jobs through the workflow API and the interactive browser sandbox.

// why local

Professional-grade research tools, without the $24,000 bill.

ELSEWHERE$24,000 a year for a terminal subscription.
VS
FNTK PRO LOCALFree and MIT licensed. Every line auditable.
ELSEWHERECloud AI APIs that log every question you ask.
VS
FNTK PRO LOCALBound to 127.0.0.1. Your queries stay on your machine.
ELSEWHEREClosed, black-box models.
VS
FNTK PRO LOCALQwen3-14B on your GPU. Every weight on your own disk.

Your research process is valuable. FNTK Pro Local keeps it private — every answer, embedding, and memo is computed on your own hardware.

// install

Up and running in 4 commands

your-machine — bash
$ pip install git+https://gitlab.com/standardhorizons/fntkprolocal $ pip install torch --index-url https://download.pytorch.org/whl/cu128 # one-time setup: 2 API keys, downloads Qwen3-14B (~8.4 GB) $ fntkpro setup $ fntkpro serve $
Then open http://localhost:8000 — the Sandbox page lets you fire every endpoint from the browser. Full docs at /api-docs.
Linux / macOS: grab llama.cpp yourself (brew install llama.cpp) or point at any OpenAI-compatible server you already run: fntkpro setup --llm-url http://127.0.0.1:11434/v1
One data key, that's it. A sec-api.io key (free tier available) pulls the raw filings. Everything else — every answer, embedding, and memo — is computed on your hardware.
// the build sheet

What's under the hood

BrainQwen3-14B GGUF8.4 GB, auto-downloaded on setup
Embeddingsbge-m32.3 GB, local vector search
Enginellama.cpp CUDA serverprebuilt binary on Windows
VRAM~16 GBsmaller GGUFs run on less; CPU-only works, slowly
Swap-inOllama / LM Studio / vLLMany OpenAI-compatible server
Disk~12 GB + SQLite cachefilings parsed once, cached 120 days
External calls1 data providersec-api.io — nothing else leaves
Binds to127.0.0.1no auth needed — it never faces the internet
// no surprises

The fine print, printed big

Patience is a position

First question per ticker downloads and parses that filing (cached after). Expect ~30–90s per answer on a 16 GB GPU. Good research takes a minute. Literally.

No GPU? No problem

Smaller GGUFs run on less VRAM, CPU-only works (slowly), and any OpenAI-compatible server slots in. The garage scales to whatever you've got.

Honest 501s

A few endpoints return 501 — they need data pipelines that aren't in this repo (fund databases, PACER, live web search). It's MIT. PRs welcome.

Company discovery corpus

Want "find similar companies"? Run fntkpro build-corpus — resumable, GPU-heavy, largest companies first. Queries get generated and cached on the spot.