Kev
jaredpalmer/kevReviewed by AI
What it does
Kev is a family of small AI models that read a piece of text and answer set questions about it: yes or no, pick one option, or give a rating, each with a probability. It is for developers who classify or route content, such as support tickets, and want it running on their own machine or server instead of a paid API. The server copies TypeSafe's System One API, so code written for Jev can point at Kev. You need Python 3.12 or 3.13, the uv tool and the terminal; a GPU for speed. A browser demo exists for the two smallest models.
How easy it is to install
A · download and open · Web
- macOS C
- Linux C
- Web A
The best grade is on the web: the README links a Hugging Face Space that runs the two smaller models with nothing to install. Running it yourself is grade C on Mac and Linux, because you install Python 3.12 or 3.13 plus the uv tool, clone the repository and run a few terminal commands; the bigger models need a large GPU (Kev-27B asks for an 80 GB card or a 96-128 GB Mac), and Windows is not mentioned in the instructions.
Numbers on GitHub
- Stars
- 9,737on Oct 31, 2026
- Downloads
- 1,09482 this week
- Growth
- +166stars in 5 days
Stars over time
Facts
- Runs on
- Linux, macOS, Web
- You need to install
- Nothing extra
- To build from source
- Python, Graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- Modal (only to deploy a hosted endpoint), Vercel AI Gateway (only to compare against the Jev service)
- Cost
- Free · The code and the model weights are free. Running it costs hardware: the README quotes rented GPUs from $0.80 to $6.25 per hour and about $1 for one fine-tuning run.
- License
- Apache-2.0 (Permissive)
- Can replace
- Jev
- Last update
- Oct 22, 2026
Risk signals (5)
- AttentionVery new repository with many stars
9621 stars in 40 days
automatic check
- Attentionpackage name collision
README.md warns: "PyPI name collision: `pip install kev` installs a **different package** — K.E.V. ORM ... This project is **not** published to PyPI. Install from source as below."
AI review
- Infoserver without key by default
README.md: "The server binds to 127.0.0.1 (`--host 0.0.0.0` to accept other machines) and is open by default; set `KEV_API_KEY` to require `Authorization: Bearer <key>` on `/v1/*`."
AI review
- Infodownloads code at setup
The deploy instructions fetch and run a script from the internet: "curl -LO https://raw.githubusercontent.com/jaredpalmer/kev/main/skills/kev-deploy/scripts/kev_serve.py" followed by "modal deploy kev_serve.py"; the fine-tune path uses "npx skills add jaredpalmer/kev@kev-finetune".
AI review
- Infoaccuracy claims self measured
The README's comparison tables against the Jev service were produced by the project itself and it says so: "We don't know what Jev was trained on, so this isn't a controlled comparison of the two architectures."
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether the published accuracy and speed numbers reproduce on your own data and hardware
- Real disk use after the model files download (the 0.8B adapter is small, the 27B model is 51 GB of weights)
- Whether it runs on Windows, which the instructions do not cover
- The real monthly cost of a hosted endpoint on Modal, including the warm container option that keeps a paid GPU running while idle
- Whether the Hugging Face browser demo keeps or logs the text you paste into it