Ollaya – Ollama for open-source, Jev-style decision models

Original Article ↗ Hacker News Discussion ↗

Open-source developers are racing to replicate Jev-style “system one” decision models — fast, cheap classifiers that can replace many LLM-based judgment calls — with projects like Ollaya and Laya aiming to run these models locally via an Ollama-like interface. Commenters debate how novel Jev really is versus classic classifiers, whether “decision model” is mostly marketing, and how close open models like Decider or Kev come to Jev’s accuracy and calibration. The thread also highlights practical use cases (moderation, smart grep, automation), the trade-off between speed and decision quality, and what rapid open-source copying means for AI startups’ moats and business models.

Project overview & positioning

  • Ollaya lets users run Jev-style “decision models” locally with a Jev-compatible API.
  • Several commenters see it as an “Ollama for decision models,” noting that Ollama or vLLM could eventually support the same endpoints.
  • The project is praised for making open models easier to run and for API compatibility that reduces vendor lock‑in.

Are “decision models” just classifiers?

  • Many argue that “decision model” is marketing for multi-class, zero‑shot text classification:
    • Input text is encoded with a transformer; output is probabilities over a fixed decision set.
    • Some call “system one models,” “decision models,” “noul,” “confidence,” etc. mostly branding.
  • Others counter that general‑purpose, well‑calibrated, multi‑task zero‑shot classification at Jev’s speed/price is non‑trivial and practically useful, even if architecturally simple.

Model quality: Jev vs Laya and other open models

  • Multiple reports say Laya is significantly weaker than Jev, especially on harder or more varied queries; it’s fast and small but trades off accuracy and stability.
  • Benchmarks shared in the thread place Jev near the top, with some open decider models and larger transformers (e.g., diffusion‑based, Qwen‑based) close behind; Laya ranks much lower.
  • Observers note a big gap between public and private benchmark scores, raising questions about generalization.
  • Some say current Jev‑like open models must be relatively large to approach Jev’s performance.

Use cases and practical value

  • Suggested uses: support ticket tagging, moderation, “smart grep” on logs, semantic shell history, home automation intent detection, and fast agent “action selection.”
  • Several developers report strong cost and latency savings vs. using LLMs for the same classification, and describe successful integrations into products.
  • Others struggle to find compelling use cases or criticize example tasks (e.g., refund vs. churn risk) as poorly specified.

Open source, competition, and moats

  • Debate over whether Jev has any real moat if similar open models appear within weeks; many think the real moat is training data and RL/synthetic data pipelines, not architecture.
  • Some see Jev as over‑marketed, others as a legitimate but incremental innovation whose main value is packaging and convenience.
  • Broader concerns surface about hype, grift, and how quickly OSS can copy commercial ideas.

Tooling and ecosystem

  • Commenters mention upcoming vLLM support for System‑One‑style endpoints and existing gateways that already implement Jev‑compatible APIs.
  • Some note alternative inference engines and web UIs built around Ollaya and Laya, and discuss GPU/CUDA version support.