Lokendra Singh.
← ALL WRITINGS
// Writing · Sep 26, 2026 · 1 min read

Jev vs Laya: two System One models

llms

Jev launched on Sept 15 with a $40M seed. Three days later, an Indian founder open-sourced Laya and called it 7.8x faster. I read both model cards.

Both are "System One" models. No text output: you send state plus typed questions (pick, score, yes/no) and get calibrated probabilities in one pass.

Jev - TypeSafe AI, San Francisco

  • Closed API, early access
  • $0.042 per 1M input tokens, output free
  • 70–500 ms end to end (TypeSafe's figure)
  • Up to 255 options per choice

Laya - Convai Innovations, India

  • Open weights, Apache 2.0
  • 421M English + 322M multilingual, 100+ languages
  • 32.8 ms per question on a Tesla T4
  • Same request shape as Jev - swap the base URL

The honest part

  • "7.8x" compares a local GPU pass with Jev's hosted API timed over the network (~236–276 ms). Not a clean race
  • Laya's 0.766 accuracy is fine-tuned on the benchmark's own training split. Zero-shot: 0.362, below the 0.461 majority baseline
  • Laya ships over-confident until you calibrate it on your own data
  • Jev wins big label sets - Banking77 0.870 vs 0.425
  • Jev's "193.6x faster" comes from TypeSafe's own workflows

My take

Jev is a product, Laya is a starting point. Zero-shot today, no GPU → Jev. Data stays on-prem, labels to fine-tune on → Laya. The real win is the pattern: stop asking a chat LLM for JSON on a yes/no.

Which decision in your stack is still an LLM call that shouldn't be? 👇

Sources

// Originally posted on LinkedIn ↗