Agent tool-routing encoder trained for just $6.60 of A100 compute released on Hugging Face
MaziyarPanahi · x · 2026-10-02
MaziyarPanahi released ModernJEV-Decide-Preview on Hugging Face, a small encoder that decides an agent's next move:
- Architecture: a 149.6M-parameter ModernBERT cross-encoder with a scalar scoring head, 4096-token inputs. Choice labels and descriptions are input text, so ranking isn't limited to a fixed tool vocabulary.
- Cost: the 60,000-decision prototype cost about $6.60 in A100 compute including setup and evaluation ($5.40 for training itself).
- How it works: it reads the conversation, policy, and available tools, scores the supplied choices, and the application executes the top-ranked action.
- Evaluation: the final checkpoint handled all 60,000 selected training decisions; held-out results cover 1,158 next-action, 542 tool-selection, and 3,652 unseen When2Call decisions, reported separately.
- Apache-2.0 licensed; explicitly an experimental preview—a Jev-style choice model, not a reproduction.
Related event: Developer Trains Open-Source Tool-Routing Model for $6.60(4 posts)→
More from coding & agent
- Handy Claude Code mod renders Mermaid diagrams right in the terminal — teropa · 2026-10-02
- Waysway Launches Travel MCP Connector for Hotels, Flights and Restaurants — modelcontextprotocol · 2026-10-02
- Desearch MCP Server Brings Real-Time AI Search on X and the Web to Agents — modelcontextprotocol · 2026-10-02
- Dev lets Sol Ultra improve his engine for 2 hours — usage barely moved 2% — Dimillian · 2026-10-02
- Meme: Developers hitting 'Allow' in Claude Code over and over — aishashok14 · 2026-10-02
- Microsoft's Autopilot isn't Scout 2.0: a private-preview shift to a Foundry-hosted cloud agent platform — pswider · 2026-10-02