Jeff releases Qwen3.5 and Gemma 4 fine-tunes for zero-shot classification.

An independent development team has released Jeff, an open-source framework designed to deploy lightweight, high-speed decision models for zero-shot classification. Built on fine-tuned iterations of Qwen3.5 and Gemma 4, the project enables developers to integrate calibrated, single-pass classification directly into local applications without relying on cloud APIs or large language models. Jeff operates by accepting plain-text scenario descriptions alongside predefined options, returning a single forward pass that outputs calibrated probabilities for each choice. The initial release features 0.8 billion and 2 billion parameter variants based on Qwen3.5, alongside a Gemma 4 E2B configuration. Unlike traditional generative models, Jeff produces no intermediate text, eliminating parsing overhead and delivering decisions in approximately 22 to 29 milliseconds on NVIDIA RTX PRO 6000 workstations, and 28 to 60 milliseconds on Apple Silicon. The training pipeline emphasizes complete local deployment. Synthetic training data was generated on-premise using the open Qwen3.8-Flash-Next model, with a dedicated leak filter to preserve dataset integrity. All fine-tuning, weight updates, and temperature calibration steps occur on local hardware, ensuring no closed-model outputs contaminate the training process. The framework supports three distinct classification formats: standard multiple-choice with up to 255 options, binary probability outputs, and continuous scoring scales. Benchmark evaluations demonstrate that Jeff architecture approaches or exceeds the performance of significantly larger models in classification and grounding tasks. On aggregate metrics across five public datasets, the fine-tuned Qwen3.5-2B variant achieved an 83.1 overall score, closely tracking Jev published 83.0 and remaining competitive against the larger AutoJev-27B model on specific benchmarks. While reasoning-heavy evaluations like Big-Bench Hard remain below larger competitors, the project zero-shot capabilities are notable. In unstructured gameplay tests involving Doom, Frogger, and Pac-Man, Jeff successfully navigated game states without prior task-specific training, outperforming random baselines and matching hand-coded rule engines in several categories. For developers, Jeff prioritizes rapid domain adaptation. A short fine-tuning cycle on proprietary examples can dramatically improve accuracy, with one test case showing held-out performance jumping from 31.7 percent to 95.8 percent in under thirty minutes. The architecture natively supports batched independent queries, allowing multiple classification tasks to be resolved in a single request. Code is released under the MIT license, with model weights distributed under Apache 2.0. The project explicitly notes its architectural lineage from the AutoJev recipe but operates as an independent initiative, offering a streamlined, resource-efficient alternative for latency-sensitive local AI deployments. This release underscores a growing industry shift toward specialized, edge-optimized neural networks that prioritize deterministic decision-making over generative complexity, enabling faster, more predictable AI integration across enterprise and developer workflows.