chenz53/Jev-alpha-26B-A4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jev-alpha-26B-A4B by chenz53 is a 26 billion parameter standalone decision model derived from Google Gemma 4. It is specifically fine-tuned to read a state and typed questions, then expose probabilities over allowed options, making it ideal for structured decision-making tasks. This model retains the Gemma4ForConditionalGeneration architecture and tokenizer, with its primary differentiator being its specialized text decision behavior, achieving 79.50% accuracy on 1,100 questions across various datasets. Its main use case is providing probabilistic outputs for defined choices rather than general chat or other modalities.

Loading preview...

Overview

Jev-alpha-26B-A4B is a 26 billion parameter decision model developed by chenz53, derived from Google Gemma 4. It is specifically designed to process a given state and typed questions, outputting probabilities for a set of predefined options. This model comes with full merged BF16 weights, eliminating the need for a LoRA adapter during inference. While it maintains the upstream Gemma4ForConditionalGeneration architecture and tokenizer, its fine-tuning focuses exclusively on text decision behavior, with other modalities and general chat capabilities not being validated.

Key Capabilities

  • Specialized Decision-Making: Optimized to provide probabilistic outputs for structured choices based on input state and questions.
  • Standalone Operation: Contains merged BF16 weights, allowing direct use without external adapters.
  • High Accuracy: Achieves 79.50% accuracy on 1,100 questions across diverse datasets like CLINC150, BoolQ, MMLU, and SciQ.
  • Calibrated Outputs: Includes a fitted temperature (1.736507) for improved probability calibration.
  • Flexible Question Types: Supports 'choice', 'noul' (no/yes), and 'score' question types with 2-26 options per question.

Good for

  • Applications requiring a model to make structured decisions and output probabilities over defined options.
  • Systems needing to route queries, classify intents, or score options based on textual input.
  • Developers looking for a specialized model for decision-making tasks where general chat capabilities are not required.

Finetunes
1 models