Unknown · 35× · vu le 02 oct. 2026
100
Momentum
Jev is the first "System One" model from TypeSafe AI, released on September 15, 2026 in Early Access. Unlike classical LLMs, Jev does not generate streaming text; instead, it accepts unstructured "State" plus typed questions (Choice, Score, Noul/Boolean) and returns directly calibrated probabilities and confidence values—without Token-for-Token generation. The model was developed by Diogo Almeida (co-developer of ChatGPT/RLHF at OpenAI) and his startup TypeSafe AI and is available on Vercel AI Gateway, Cloudflare, LangChain, and other platforms. According to the provider, Jev is up to 194–200x faster and up to 400–445x cheaper than comparable language models for classification and routing tasks.
Historique du momentum
04.07.02.10.
Fonctionnalités
| Throughput/Latency | End-to-end response time 70–500 ms (typically ~100 ms); rate limit ~250,000 tokens/sec and 1,200 requests/min (early access, dynamic) |
| License | Proprietary model, served through TypeSafe's API and licensed cloud platforms (not open source) |
| Platform | Own API (console.typesafe.ai), plus Vercel AI Gateway, Cloudflare Workers AI, OpenRouter, Requesty, LangChain, Langfuse |
| Price | $0.042 per 1M input tokens; output tokens free |
| Protocol Compatibility | Custom request format (state/questions, /v1/systemone), not standard OpenAI chat-completions format; wrapped OpenAI-compatible via some gateways (e.g. Requesty) |
| Release Date | September 15, 2026 (early access), version 1.13 on September 18, 2026 |
| Supported Models/Providers | TypeSafe AI (primary provider); also accessible via Vercel AI Gateway, Cloudflare, OpenRouter, Requesty, Impossibl |