Jev Hits 36M Views in 2 Days, Community Ships 6 Open Clones

Latent Space · rss · 2026-09-19

Latent Space's AINews spotlights the "Jev wave": the non-generative decision model's launch video hit 36M views in two days (vs OpenAI's Navier Stokes at 74M and Anthropic's Fable 5 at 57M), and being closed-source fueled six community clones. Best guesses put its architecture near ModernBERT encoders plus Diffusion.

Open clones

Laya: 421M params, ModernBERT-large + two scoring transformer layers, PPO over sequence embeddings
Bespoke Nimble: LoRA finetune of Qwen3.5-9B with synthetic contrastive data; 66%→90% vs Jev's 93%, 100ms on H100
SemIf (fka OpenJev): Qwen3.5 4B/35B backbone with a three-class NLI head
Kev-0.5B: Qwen2.5-0.5B LoRA adapter, runs on a MacBook Pro
Training data is acknowledged to be 100% synthetic; no standard benchmark exists for this category

Also in this issue

Jev is framed as a near-zero-marginal-cost System 1 layer (400x cheaper scoring) for routing, citation selection and escalation; LangChain's Harrison Chase called browser use its best application, and Cline shipped a Jev plugin
Claude Code v2.1.277 now reads AGENTS.md when CLAUDE.md is absent, cementing it as a cross-tool standard
The "Harness Tax" analysis shows a minimal tool set (read/write/edit/bash) can reach the Pareto frontier while cutting spend
Model choice is bifurcating: frontier models for planning, cheap ones for execution — one team cut spend 2 orders of magnitude moving from Opus to GLM Flash
GPT-6 Astra solved another FrontierMath open problem; GPT-6-astra leads RSI-Exam at 0.5126

Original post →

Jev Hits 36M Views in 2 Days, Community Ships 6 Open Clones

More from coding & agent