System 1

The practical criterion for choosing a model is to count how many steps the answer requires: use System 1 for one-step answers and System 2 for multi-step answers.

Jev's design borrows the System 1 concept from psychology, serving as a fast intuitive judgment role in the process.

System 1/System 2 theory originates from Daniel Kahneman's Thinking, Fast and Slow: System 1 is fast, automatic, intuitive judgment, and System 2 is slow, deliberate reasoning that requires working memory.

Zach proposed that when changes in confirmation copy cause strict text assertions to misjudge, a System 1 model (such as Cloudflare's Cleft) can be introduced to make a fast semantic judgment of whether the result matches the intent.

In AI Agent architectures, as many as 80% of judgment nodes only need System 1-style intuitive reflexes rather than token-by-token autoregressive generation by a language model.

System 1 targets decisions over given options rather than open-ended answer generation, and the upper limit for selection tasks is 256 options.

System 1 does not need to generate an answer and then parse it; probabilities are read directly from a single forward pass, which is more practical than free-text commentary for scenarios requiring stable numerical values, such as offline evaluation, filtering, and reward signals.

JEV-27B-VL exposes two distinct interfaces: System 1 reads logits and returns typed decisions and calibrated probabilities, supporting yes/no judgments, selection among 2 to 256 options, and scoring from 0 to 5, with prompts up to 256K tokens; System 2 is described as unmodified Qwen3.8-27B, which accepts image input and outputs text or step-by-step reasoning.

For tasks requiring multi-step reasoning, building a System 2 scaffold for a System 1 model at the application layer (providing scratch space, introducing external tools for verification, setting checkpoints, allowing multiple attempts and picking the best) is often more effective than simply switching to a larger model.

Jev is positioned by its developer TypeSafe as the first System 1 model, focused on classification decisions rather than text generation.

The decision model evaluates state against bounded criteria and returns a decision score without producing an open-ended token stream, and runs in milliseconds.

The recommended collaborative system design is: use a System 1 one-glance model to answer all questions while retaining confidence scores, and hand off tasks below a threshold to a System 2 thinking model.

At the model level, System 1 corresponds to a single forward pass, where the input traverses the fixed layer stack only once.

Each forward pass of the original Transformer has a fixed computational depth that cannot be dynamically extended, making it naturally correspond to System 1.

For tasks that can be answered in one step, such as classification and judgment tasks, the one-shot direct output of System 1 is the optimal solution, and using System 2 only adds unnecessary latency and cost.

When reasoning chains grow longer and multi-level derivation is required, the accuracy of System 1 models collapses sharply, because a single forward pass cannot carry too many intermediate states.

A System 1 model performs only a single forward pass when faced with a question and directly outputs the answer, with no intermediate drafts.

The naming of the /v1/systemone interface echoes the concept of "System 1" in psychology — intuitive, fast, and automatic thinking.

Reasoning models have introduced mechanisms such as chain-of-thought, reflecting the evolution of AI from System 1 fast intuitive responses toward System 2 slow deep thinking.

TypeSafe (whose parent company is System 1) has launched an AI model called Jev, which does not generate text but instead produces decisions for code execution.