Laya: an open 421M decision model that classifies in 33ms
Convai's ModernBERT head is 7.8x faster than TypeSafe's Jev and free to self-host, if you fine-tune it. Plus: ECC for Claude Code, OpenAI key lockdowns.
Copy markdown
A 421M head that decides in 33ms
Convai Innovations dropped Laya, an Apache-2.0 decision model built on ModernBERT-large that takes an email, JSON, or a trace and returns a typed answer with a confidence score in one forward pass — about 33ms on a T4, ~7ms batched. No autoregression, no tokens streamed.
It undercuts TypeSafe's Jev by 7.8x
Laya is the open answer to Jev, TypeSafe's closed, metered decision API that runs ~150-276ms per call. Same non-autoregressive idea, but Laya self-hosts under Apache-2.0 at 32.8ms median — 7.8x faster — so a routing or moderation step stops being a per-call API bill.
The catch: fine-tune it, don't trust it cold
The model card is blunt: Laya is "a fast base to specialise, not a zero-shot decision engine." Cold, it barely beats random (0.362); fine-tuned it hits 0.993 on spam and 0.980 on phishing. Refit the temperature before you believe a confidence score, and keep options under ~20.
Build this weekend: kill your classifier bill
Point Laya at the yes/no work you're currently paying an LLM to do — invoice sorting, ticket routing, guardrails, agent-trace tagging. The 322M multilingual checkpoint covers 100+ languages at 1024 tokens. `pip install laya`, fine-tune on a few hundred labels, run it local.
Elsewhere: ECC hardens Claude Code and Cursor
affaan-m/ECC is trending (+876 stars in 24h): an open "agent harness" that bolts skills, memory, and security onto Claude Code, Codex, Opencode, and Cursor. If you live in one of those daily, it's a drop-in layer worth a look this week.
Elsewhere: OpenAI lets orgs lock down API keys
OpenAI's platform now lets admins restrict who can mint API keys — service-account-only, project-owned-only, or no new keys at all — and enforce max key lifetimes. If you're on a team account, expect key hygiene rules to tighten; existing keys keep working.