PostHog/jeeves

Jeeves – Reasoning improves Jev-like decision models

What it solves

Jeeves addresses the limitation of "Jev-style" decision models, which are often used for calibrated classification tasks but suffer from low accuracy. By incorporating a reasoning step before making a final decision, Jeeves improves performance on complex tasks and out-of-domain data compared to existing models like Jev and Kev.

How it works

Jeeves is a 9B parameter model based on Qwen3.5 that uses a combination of Supervised Fine-Tuning (SFT) and CISPO (a training method for scaling test-time compute). It uses a "pointer head" to score options based on a reasoning chain generated before the final decision. The system includes a diffusion-based "drafter" to speed up inference and supports FP8 quantization to reduce memory usage on both NVIDIA GPUs and Apple Silicon.

Who it’s for

This project is for developers and researchers who need a high-performance, calibrated decision-making engine for classification tasks (e.g., categorizing customer support tickets or scoring sentiment) and want to experiment with test-time reasoning to improve accuracy.

Highlights

  • Outperforms Jev and Kev-9B on JevBench and various held-out test datasets.
  • Supports multiple question types including yes/no, multiple-choice, and scoring within a single request.
  • Includes a drop-in Python SDK compatible with existing Jev-style implementations.
  • Features a diffusion-based drafter that significantly accelerates inference speed.
  • Provides full training, evaluation, and export pipelines, including support for FP8 precision.

Related

  • Dispatch
    Jeeves: Improving Decision Models with Reasoning and Diffusion Drafting

    Jeeves is a 9B reasoning classifier based on Qwen3.5-9B that outperforms Jev and Kev-9B on out-of-domain tasks by integrating a reasoning chain before decision-making.

  • Dispatch
    Kev: Open-Source Decision Models built on Qwen3.5

    Kev is a family of small, open-weight decision models (0.8B, 4B, and 9B) built on Qwen3.5 that implement a Jev-like architecture for high-speed, structured classification and rating tasks.

  • Project
    TianyuCodings/JevHarness

    JevHarness is a framework that lets developers offload repetitive decision-making to a fast judgment engine (Jev) while using an LLM to author and optimize the underlying task-specific logic.

  • Project
    jaredpalmer/kev

    Kev is a family of small, specialized decision models that provide calibrated, structured outputs for classification and scoring tasks, designed as an open-source, self-hostable alternative to proprietary decision-model APIs.

  • Dispatch
    Jeff: Jev-compatible 0.8B and 2B Decision Models for Zero-Shot Classification

    Jeff is a set of small, open-weight fine-tunes of Qwen3.5 and Gemma 4 designed for high-speed, zero-shot classification with a Jev-compatible API, achieving decision latencies as low as 22ms.