Diogo Almeida S Ai Lab · 6× · naposledy 17. 9. 2026
Jev is the first public model from TypeSafe AI, a 2024 startup founded by former OpenAI researcher and ChatGPT/RLHF co-developer Diogo Almeida. It's a "System-One Model": rather than generating text autoregressively Token by Token, Jev accepts an unstructured state plus typed queries and delivers typed, calibrated probability decisions (selection, score, yes/no) in a single parallel pass. It was trained using a new method called Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest calibration rather than human-preferred answers. According to the manufacturer, Jev achieves comparable intelligence to existing large language models on decision tasks ("System-One Tasks").
Vlastnosti
| Key-Benchmark (%) | 67,8% Übereinstimmung im TypeSafe-4-Workflow-Eval (vergleichbar mit GPT-5.6 Terra: 67,9%; GPT-5.6 Sol: 74,1%; Opus 5: 73,1%) |
| Kontextfenster (Token) | ca. 32.000 Token (State + Fragen zusammen), entspricht ca. 150.000 Zeichen englischem Text |
| Lizenz | Proprietär, geschlossenes Modell, ausschließlich über TypeSafe-API nutzbar |
| Multimodalität | Keine Bild- oder Audio-Eingabe zum Launch; nur strukturierter Text-/JSON-State plus typisierte Fragen |
| Plattform | Gehostete API von TypeSafe AI, Early-Access-Warteliste, Browser-Playground; kein On-Prem/Open-Weight-Zugang |
| Preis pro 1M Token | $0,042 pro 1M Input-Token; Output-Token kostenlos (nicht bepreist) |
| Release-Datum | 15. September 2026 (Early-Access-Start) |