TypeSafe launches Jev: a decision-only model 200x faster and 400x cheaper than frontier LLMs

Latent Space · rss · 2026-09-16

Latent Space's AI News covers TypeSafe's Jev, a "System One" decision model trained with RLCD that only classifies/routes/scores — 20–200x faster and 40–400x cheaper than small frontier LLMs with free output tokens; community frames it as a cheap calibrated inference engine for structured choices, pairing well with DSPy-style signatures.

Also: Periodic Labs' Neon, a 1T-param materials-science model trained on 1300 H200s in a lab-RL loop (Kimi K2.x base, FrontierXRD 2.7%→55.3%, beating GPT-6 Astra and Claude Fable 5.1); Google's Gemini 3.8 Live (97 languages, async tool calls, #1 speech-to-speech at 82.6, $0.84–$3.50/hour pricing); Devin gaining native Mac VMs across a Rust-rebuilt cloud agent platform; LangChain shipping MCP servers for managed agents; and a Microsoft paper showing plain bash beats typed tool catalogs by 21.8–24.5 points on TheAgentCompany.

Original post →

More from Models

Models channel →