Engines: Kenning, Clef, Jev

Several engines speak the same System One wire format (POST /v1/systemone). The systemone client works with all of them, so the code on this site runs unchanged on each. You switch engines by changing a URL.

	Kenning	Clef	Jev
Made by	the SystemOne Builder community	Cloudflare	TypeSafe AI
Licence	Apache-2.0 weights	Apache-2.0 weights	hosted service, paid
Runs	on your machine	on your machine	on TypeSafe’s API
Size / hardware	435M parameters, ~2 GB of GPU memory, or CPU	9B parameters, a 24 GB GPU	n/a
Determinism	same request, same answer (same hardware)	yes, locally	per TypeSafe
You can train it further	yes: SystemOne Builder	as a base, yourself	no

Same benchmarks, side by side

We run every engine through the same suites with systemone bench (in SystemOne Builder). None of these suites were used for training. Numbers are accuracy unless noted. Kenning is kenning-large-v0.4; Clef is clef-flash on one RTX 3090; Jev is jev-latest (October 2026).

Suite	Kenning	Clef	Jev
Modern emails (20 held out, incl. calm credential lures)	0.75	0.90	0.90
…phishing emails it was sure were safe (auto-closed)	0	0	0
Phishing dataset (50)	0.78	0.96	0.96
Out of domain: spam / emotion / news topic	0.917 / 0.583 / 0.867	0.900 / 0.600 / 0.950	0.967 / 0.583 / 0.933
Latency per request (p50)	33–88 ms	220–290 ms	~150 ms incl. network

How to read this honestly:

Kenning is smaller and faster, and behind on subtle phishing. It’s about 20 times smaller than Clef. What matters most is that in every suite it was right whenever it was sure, apart from one spam message. It stays humble rather than confidently wrong, and sends more items to a person.
Twenty or fifty items is a small sample. Each difference of one or two items moves these numbers by several points. The suites are open: run them yourself, and add your own.
Full method, per-suite details and caveats are in the Kenning docs.

Pointing the client at each engine

import os
from systemone import Client, Kenning

# Kenning in-process (pip install "systemone-client[local]")
model = Kenning.from_pretrained("systemonedev/kenning-large-v0.4")

# Kenning or Clef served by SystemOne Builder (localhost only, no key)
kenning = Client("http://localhost:8093")
clef = Client("http://localhost:8094")            # docker compose --profile clef up -d clef

# TypeSafe Jev (your key; each request is billed under your TypeSafe agreement)
jev = Client("https://api.typesafe.ai", api_key=os.environ["TYPESAFE_API_KEY"], model="jev-latest")

All four objects take the same system_one(state=..., questions=...) call and return the same Response.

Which one should I use?

Prototyping, privacy-sensitive data, high volume, or no budget: Kenning. It’s free, runs anywhere, and you can fine-tune it on your own labelled data with SystemOne Builder.
Hardest judgement calls on your own hardware: Clef, if you have a 24 GB GPU to give it. It’s also a good teacher to distil into a smaller model.
No GPU and no operations work: Jev, as a hosted service.

You don’t have to choose once. A common setup is Kenning first, escalating the uncertain middle band to a bigger engine or to a person. See the fuzzy if-statement.