Open-source model Jev called out for not crediting Lava, its 8192-context prior work
nischay_twt · x · 2026-09-21
JFPuget highlights Lava as clear prior work and inspiration for the open-source Jev model: Lava is faster but limited to an 8192-token context. He speculates Jev fixed this by replacing ModernBert with an autoregressive encoder like a 1.7B Qwen model, and notes Jev never acknowledged Lava.
More from Models
- GLM lists cache storage as 'limited-time free' while Kimi K3 charges $3.75 per cache write — teortaxesTex · 2026-09-21
- GPT-6's 3D hype meets reality: a gorgeous handle you should never print — ngxson · 2026-09-21
- Dev tests model routing: multi-model 'team' costs as much as Astra alone and times out — kevinkern · 2026-09-21
- Diffusion makes inference look like training—the shape GPUs were built for — victor_explore · 2026-09-21
- Developer cancels $200/mo Codex subscription, going all-in on open-source models — haydendevs · 2026-09-21
- Pre-registered test of Jev's no-hallucination claim: type safety holds, calibration splits — NervousAd4440 · 2026-09-21