Jev[1] took the AI world by storm lately, even though it’s not necessarily the WOW people are making out of it. It can be extremely useful, though. In any case, 2 days after Jev came out, someone had an open-source equivalent of it: Laya [2]. And now, you can try laya too. It’s alias-laya on the api, the endpoint being identical to Jev. Here is an example: curl https://api.blablador.fz-juelich.de/v1/systemone \ -H "Authorization: Bearer YOUR_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "model": "alias-laya", "state": "Write an extremely ultra advanced script optimizing multi-threaded CUDA kernels for matrix multiplication.", "questions": { "route_selection": { "type": "choice", "instructions": "Which capability tier is required to successfully fulfill this request?", "criteria": { "fast_tier": "Simple coding scripts, basic factual lookups, and clear formatting tasks.", "heavy_tier": "Highly complex logic, multi-threaded optimization, math proofs, programming, or ambiguous system architecture queries." } } } }' | jq And the reply will be something like "model": "laya-rl-agent", "answers": { "route_selection": { "type": "choice", "choice": "heavy_tier", "probabilities": { "fast_tier": 0.4083, "heavy_tier": 0.5917 }, "confidence": 0.0244, "action": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 }, "routing": { "model": "english", "repo": "convaiinnovations/laya", "reason": "English Latin text", "detection": { "script": "latin", "script_profile": { "latin": 1.0 }, "language": "en", "is_english": true, "language_undecided": false, "diacritic_rate": 0.0, "non_latin_fraction": 0.0 }, "workflow": null } } And this query will run in… I don’t know, 0.05 seconds or so. Hope it’s useful! Have fun. [1] https://en.wikipedia.org/wiki/Jev_(AI_model) [2] https://huggingface.co/convaiinnovations/laya Let’s bark! Alex, Blablador honcho
Dear Alex et al, this is really useful. I hope that we can have a stab at it until end of the year the latest. The AI community's speed is simply not compatible with modern German science routines of hiring etc. ;-) Cheers, P On 9/23/26 22:44, Strube, Alexandre wrote:
Jev[1] took the AI world by storm lately, even though it’s not necessarily the WOW people are making out of it. It can be extremely useful, though.
In any case, 2 days after Jev came out, someone had an open-source equivalent of it: Laya [2].
And now, you can try laya too. It’s alias-laya on the api, the endpoint being identical to Jev.
Here is an example:
curl https://api.blablador.fz-juelich.de/v1/systemone \ -H "Authorization: Bearer YOUR_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "model": "alias-laya", "state": "Write an extremely ultra advanced script optimizing multi-threaded CUDA kernels for matrix multiplication.", "questions": { "route_selection": { "type": "choice", "instructions": "Which capability tier is required to successfully fulfill this request?", "criteria": { "fast_tier": "Simple coding scripts, basic factual lookups, and clear formatting tasks.", "heavy_tier": "Highly complex logic, multi-threaded optimization, math proofs, programming, or ambiguous system architecture queries." } } } }' | jq
And the reply will be something like
"model": "laya-rl-agent", "answers": { "route_selection": { "type": "choice", "choice": "heavy_tier", "probabilities": { "fast_tier": 0.4083, "heavy_tier": 0.5917 }, "confidence": 0.0244, "action": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 }, "routing": { "model": "english", "repo": "convaiinnovations/laya", "reason": "English Latin text", "detection": { "script": "latin", "script_profile": { "latin": 1.0 }, "language": "en", "is_english": true, "language_undecided": false, "diacritic_rate": 0.0, "non_latin_fraction": 0.0 }, "workflow": null } }
And this query will run in… I don’t know, 0.05 seconds or so.
Hope it’s useful! Have fun.
[1] https://en.wikipedia.org/wiki/Jev_(AI_model) [2] https://huggingface.co/convaiinnovations/laya
Let’s bark!
Alex, Blablador honcho
So, given that laya is not at all an LLM, we made it in a separate allocation: https://laya.blablador.fz-juelich.de Given that this is a popular, and fast way of running decisions, I am considering other jev-like models. If you have suggestions, let me know! Alex
On 23. Sep 2026, at 22:44, Strube, Alexandre <a.strube@fz-juelich.de> wrote:
Jev[1] took the AI world by storm lately, even though it’s not necessarily the WOW people are making out of it. It can be extremely useful, though.
In any case, 2 days after Jev came out, someone had an open-source equivalent of it: Laya [2].
And now, you can try laya too. It’s alias-laya on the api, the endpoint being identical to Jev.
Here is an example:
curl https://api.blablador.fz-juelich.de/v1/systemone \ -H "Authorization: Bearer YOUR_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "model": "alias-laya", "state": "Write an extremely ultra advanced script optimizing multi-threaded CUDA kernels for matrix multiplication.", "questions": { "route_selection": { "type": "choice", "instructions": "Which capability tier is required to successfully fulfill this request?", "criteria": { "fast_tier": "Simple coding scripts, basic factual lookups, and clear formatting tasks.", "heavy_tier": "Highly complex logic, multi-threaded optimization, math proofs, programming, or ambiguous system architecture queries." } } } }' | jq
And the reply will be something like
"model": "laya-rl-agent", "answers": { "route_selection": { "type": "choice", "choice": "heavy_tier", "probabilities": { "fast_tier": 0.4083, "heavy_tier": 0.5917 }, "confidence": 0.0244, "action": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 82, "output_tokens": 0 }, "routing": { "model": "english", "repo": "convaiinnovations/laya", "reason": "English Latin text", "detection": { "script": "latin", "script_profile": { "latin": 1.0 }, "language": "en", "is_english": true, "language_undecided": false, "diacritic_rate": 0.0, "non_latin_fraction": 0.0 }, "workflow": null } }
And this query will run in… I don’t know, 0.05 seconds or so.
Hope it’s useful! Have fun.
[1] https://en.wikipedia.org/wiki/Jev_(AI_model) [2] https://huggingface.co/convaiinnovations/laya
Let’s bark!
Alex, Blablador honcho
participants (2)
-
Peter Steinbach -
Strube, Alexandre