Unsloth’s primary documentation describes running Laya locally behind a TypeSafe-compatible endpoint. The useful part is not another speed claim: it spells out what does not transfer when you replace Jev. I read the complete guide; I did not install Unsloth, run the model or reproduce its example responses.

What the documented integration looks like

  • Enable Settings → API → Decision API → Serve requests, download the selected model, and send state plus typed questions to /v1/systemone.
  • The guide’s example asks for a support team (choice), refund intent (noul, a yes/no probability) and urgency (score) in one request.
  • Set TYPESAFE_BASE_URL to the local server and use its access token with the TypeSafe SDK. The jev-latest alias maps to the model selected in Unsloth’s settings—not Jev’s hosted weights.

Compatibility stops at the decision policy

  • Do not reuse Jev’s confidence thresholds. Unsloth explicitly says Laya computes confidence differently and recommends gating on the relevant option’s value in probabilities instead.
  • That probability still needs testing on your own labeled data. A typed output, a sharp distribution and a correct decision are different properties.
  • The guide documents a fixed token budget shared by choice options: beyond approximately 20 described options, labels are trimmed. This is a documented limitation, not a limit measured here; accepting a large option list does not mean preserving its full meaning.

Check the operational fit

  • Unsloth lists its default multilingual model at 678 MB, with 4 GB minimum RAM and a 1,024-token context; its English variant is listed at 846 MB, 5 GB RAM and 512 tokens. These are vendor specifications, not independently measured requirements.
  • CPU is the documented default; GPU execution retains the model until unload or restart, with CPU fallback on GPU memory exhaustion.
  • Unsloth claims 10–20 seconds for cold loading and subsequent requests well below one second on most CPUs. No hardware-specific benchmark is supplied here, and those timings were not reproduced.

Treat this as a migration recipe, not an accuracy comparison or a privacy audit. Replay representative Jev requests against the local model, check truncation and ambiguous cases, and retune the action policy before letting the replacement change real records.