ENPT

Swarms 16 adds a DecisionModel client for Jev and Clef

Last updated:

Source · Model card: Jev

The Swarms multi-agent framework shipped version 16, which it calls “Overclock” (16.0.0 on PyPI on 2 Oct 2026, 16.0.1 on 3 Oct). Its headline addition for this category is a DecisionModel client, explained in a guide published 3 Oct.

What it does

  • One interface, two providers. The model name picks the backend: names starting with jev (jev-latest, jev-preview, jev-1.13.0) go to TypeSafe, and clef / clef-flash go to Cloudflare Workers AI. You only need keys for the providers you call.
  • Ensembles and cascades. The guide shows averaging several models and a cascade where Jev answers first, Clef is consulted below 0.75 confidence, and disagreements are flagged for review.
  • Inside agents. A decision model can sit in front of an LLM agent as a router, screener or judge.
  • Swarms API routes. GET /v1/decision-model/models, POST /v1/decision-model/completions, and POST /v1/systemone, which accepts TypeSafe’s exact request and response format so the TypeSafe SDKs work by changing the base URL and key.

Numbers to read carefully

A companion post, Reduce LLM Costs 10x, reports that screening with jev-latest cost 10.1× less per item than the cheapest LLM screener the team wrote and answered about 3× faster. That is a single in-house run, so it is an UnverifiedClaim, and the post says Clef was not part of the measurement. Prices in the guide (Jev $0.042 per million input tokens; Clef $0.24 and Clef-flash $0.09) come from Swarms’ built-in price table.

Why it matters

Swarms treats decision models as a provider-neutral type, not a single-vendor integration, so switching between Jev and Clef is a model-name change. It also adds another server that speaks /v1/systemone, next to Ollama and the open model servers.