Jev in 25 lines of Python: a decision model you don't have to buy

A viral repro shows Jev's trick is just logits plus softmax; Open-Jev-27B nearly matches the paid model and Laya decides in 33ms — all open weights.

Nowline SEP 24 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • What Jev is — and 25 lines that do it

    Jev, TypeSafe's "System 1" decision model, takes a prompt plus options and returns calibrated probabilities instead of generated text. A Sept 22 post reproduces the core in 25 lines: load a tiny GGUF with llama-cpp-python, read the final-token logits, then logaddexp and softmax over your choices. On an email test it returned phishing 0.885, spam 0.084, legit 0.031 — the whole trick, no proprietary API, running locally.

  • Open-Jev-27B lands one hard answer behind the paid model

    Open-Jev-27B-v1.1 (Qwen3.8-27B base, LoRA rank-8 decision head, MIT code) scores 85.28% on JevBench — 197 of 231 — trailing proprietary Jev by a single hard question, and handles Choice, yes/no, and Score-rubric decisions. You can drop a near-parity, typed-decision layer into your agent's routing or guardrails without a per-call bill.

  • Laya: 421M params, 33ms, and it can't hallucinate

    Convai's Laya (ModernBERT-large, Apache 2.0, 421M) answers typed questions in ~33ms in one forward pass with far tighter calibration (ECE 0.081 vs 0.246) and never emits text. A 322M multilingual variant covers 100+ languages, and a built-in router picks the checkpoint per request. It's a CPU-friendly triage or guardrail classifier that runs 6-8x faster than asking an LLM to label.

  • AWS open-sources Strands Harness, ~28% cheaper agent loops

    AWS's Strands team shipped Strands Harness (Apache 2.0, Python and TypeScript): a ready-made agent loop with tools, memory, and recovery behind one create_harness() call across Bedrock, Anthropic, OpenAI, Google, Ollama, or LiteLLM. It reports ~28% lower token cost than rival harnesses at near-equal accuracy over six benchmarks; on Terminal-Bench 2.1 it ran $56.29 at 69.7% vs Claude Code's $248.05 at 61.8%.