The open-source answer to proprietary system-1 models just shipped, and it's honest about its…
the open-source answer to proprietary system-1 models just shipped, and it's honest about its limits. convai's laya, a 421m non-autoregressive decision model, replies in 33ms and runs up to 8x faster than typesafe's jev, but the model card admits 0.362 zero-shot accuracy on typed decisions.
speed is not a substitute for judgment.
Context
The Hugging Face model card for convaiinnovations/laya, Apache 2.0, describes a ModernBERT-large backbone of 395M plus a decision head for 421M total, with multi-question batch latency of about 33 to 38 ms on GPU. Its table, against TypeSafe's Jev published figures, says about 10.4x faster on average; the project site says 32.8 ms and 6 to 8 times faster. The typed decisions card lists 0.362 for the base laya checkpoint, not fine-tuned, and 0.766 for the fine-tuned checkpoint against Jev's published 0.727, and the main card lists a zero-shot macro accuracy of 0.651 on held-out task families. Limitations: English only, 512 tokens per question, calibration measured on benchmark data, and arithmetic and multi-hop should stay in deterministic code.
The 0.362 belongs to the non-fine-tuned base row, so attaching it to the typed decision model, as the note does, is not supported as worded; the figures are not merged. The speed claims are the vendor's, compared with Jev's published numbers and not shown on matched hardware, and the multiples differ (4x, 6 to 8x, 7.8x and 10.4x). Speed is not a substitute for judgment is the author's line.
TypeSafe's Jev post of 15 September 2026 describes Jev as early access, hosted, with claims of two orders of magnitude faster than LLMs and response times of 70 to 500 ms; Laya is Apache 2.0 open weights and compares itself to Jev's published numbers and not a same-hardware run. Both are vendor claims with different measurement scopes, and no replication or quality verdict was read.
Related work
- Introducing System One models and Jev (TypeSafe, 15 Sep 2026) ↗The model Laya compares itself against.
Watch next
- Independent latency and accuracy replications.
Sources
- laya (Hugging Face)huggingface.co
- laya-typed-decisions (Hugging Face)huggingface.co
- Laya project sitelaya.convaiinnovations.com
- Introducing System One models and Jev (TypeSafe, 15 Sep 2026)typesafe.ai
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 10:17 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →