Kahneman, System 1.
The name is a nod to Daniel Kahneman's System 1: fast, cheap judgments that settle most cases and route the rest before a slower System 2 (a frontier model or a person) steps in.
Kahneman, a psychologist awarded the 2002 Nobel Memorial Prize in Economic Sciences, described the two systems in Thinking, Fast and Slow (2011). System 1 is quick, automatic and intuitive; System 2 is slow, deliberate and effortful.
One state. Three typed questions.
I was charged twice for order #4821 and nobody answers. Fix this before Friday or I cancel my account.
Choice
PICK ONEWhat does the customer want?
Score
ORDERED SCALEHow urgent is this?
Noul
TRUE / FALSEThe customer threatens to leave.
Getting a decision out of text
Kahn1
OPEN · LOCALReads typed answers from option logits, each with a calibrated distribution. Open weights, runs on your own hardware, no per-call price. Production throughput needs a GPU, and thresholds should be recalibrated on your domain.
JEV
HOSTED APIA hosted API with strong accuracy and nothing to run yourself. Every call is paid, goes over the network and sends your data to a third party.
Other open-source LLMs
OPEN · LOCALGeneral-purpose models you can run locally. They generate their answer as text, so it has to be parsed, latency grows with the output and confidence is rarely calibrated.
JSON mode
ANY CHAT MODELAvailable on almost any model or API. The model writes JSON as text: you validate and parse it, malformed answers happen and no probabilities come back.
Structured outputs
ANY CHAT MODELStructured decoding constrains the tokens a model generates, and the result still has to be parsed. The output fits the schema, but there is no calibrated distribution.
Classifiers
TRAINEDBERT, XGBoost… Under 1 ms to 10 ms on a CPU and strong on tabular data. The ontology is fixed: a new category means relabel, retrain, redeploy.
Kahn1 or classical ML?
| Classical ML (BERT, XGBoost…) | Kahn1 | |
|---|---|---|
| New category | Relabel, retrain, redeploy | Edit the schema in the request |
| Labelled data | Hundreds to thousands per class | None to start; some to recalibrate |
| Latency | Under 1 ms to 10 ms, CPU | 36 ms (3B) to 89 ms (4B) median, GPU |
| Hardware | Small CPUs, edge devices | A GPU for production throughput |
| Explanations | Feature importances, tree splits | Full distributions, no feature attribution |
| Determinism | Same input, same output | Same weights, prompt and kernels give the same logits; nothing is sampled |
| Audit trail | A score or a label per decision | A full probability distribution per decision, ready to log |
Pick classical ML when the ontology is fixed for years, latency must stay under 5 ms, hardware is small or the data is tabular. Pick Kahn1 when categories change often, when you start without labels, or when the signal is in free text.
A decision you can't reproduce is a decision you can't defend.
A generative LLM samples its answer: ask twice, and the same case can get two different decisions. In the high-risk uses listed by the EU AI Act, such as credit scoring, hiring or access to essential services, providers must keep logs and allow human oversight. A decision that changes between runs is hard to document, audit or contest.
Kahn1 samples nothing. The same input gives the same logits, and every answer comes with its full distribution, so each decision can be replayed and logged as it was made. Low-confidence cases can be routed to a person by threshold.
Reading logits instead of parsing JSON removes format errors, not the obligations that come with using a neural model for decisions (EU AI Act, sector rules). Keep a person or a larger model on the low-confidence path, and log what was decided.
Use cases
Kahn1 is meant to be the fast first layer of a pipeline: it takes the routine decisions and hands the rest to something slower and more expensive.
Triage and routing
Sort tickets, emails or claims by intent, urgency and risk in one pass. To change the categories, edit the schema: no relabelling, no retraining.
Guardrails for agents
Before a tool call runs, ask yes/no questions about it: does it leak personal data, does it break a policy, did the user approve it? Each answer is a probability you can gate on.
Document scoring
Rate tone, urgency, complexity or quality as continuous expected values rather than coarse labels, and index documents on them.
Selective escalation
Automate the confident cases and send the rest to a larger model or a person. With calibrated confidence, the threshold becomes a risk decision you can measure.
Cold start
A new product line or a new taxonomy works zero-shot on day one. Recalibrate on your own labelled examples with sysone calibrate once you have them.
Control loops
One Choice per tick can drive a policy: Kahn1 plays Snake as well as a greedy heuristic, provided the state spells out what lies around the head.
Read this before real decisions
Kahn1 comes in two sizes, 4 and 3 billion parameters. Read this before putting either in front of real decisions.
- It is not a frontier model.
- Accuracy depends on the primitive.
- Calibration holds on its own distribution.
- Confidence is not the probability of being right.
- Deterministic computation is not deterministic truth.
- Regulation still applies.
- The speed comes from engineering.
- The browser demos run an older checkpoint.