finds.dev

microsoft/JARVIS

JARVIS, a system to connect LLMs with ML community. Paper: https://arxiv.org/pdf/2303.17580.pdf

View on GitHub

★ 25,548PythonMITupdated Jul 2025

// the take

JARVIS is Microsoft Research's reference implementation of HuggingGPT: an LLM (originally text-davinci-003) acts as a controller that plans tasks, picks Hugging Face models to execute them, and stitches the results into a response. It's aimed at researchers exploring LLM-orchestrated multi-model pipelines, not at anyone wanting a production agent framework.

The four-stage pipeline (plan → select model → execute → generate response) is a clean, well-documented pattern that's since been copied widely, and the repo bundles the full paper's demo/eval data so you can reproduce the original results. It also folds in two follow-up research projects (TaskBench, EasyTool) with their own datasets and benchmarks, giving it genuine research value beyond the original demo. Deployment modes are flexible — pure Hugging Face Inference Endpoints, fully local, or hybrid — so you're not forced into the heaviest setup just to try it.

The core controller is hard-wired to text-davinci-003, an OpenAI model that's been retired; getting this running today means patching the LLM call yourself, and there's no indication that's been done upstream. The 'default' local config wants 24GB VRAM, 80GB RAM, and 284GB of disk just to download the expert models, which prices out casual use. It's a frozen research artifact, not a maintained project — no work since the January 2024 EasyTool release, and the web API has no auth, rate limiting, or input validation, so it's not something to expose beyond a local demo.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

search the same ground:

// at a glance

Stars
25,548
Forks
2,270
Language
Python
License
MIT
Last commit
29 Jul 2025

// nearby

If JARVIS isn't quite it.

All repos →