groxaxo/gemma4-prometheus-merged

VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 3, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

groxaxo/gemma4-prometheus-merged is a 31 billion parameter language model based on Google's Gemma-4 architecture, specifically the `google/gemma-4-31B-it` variant. This model has been enhanced with Prometheus adversarial-steering optimization, with the steering vectors merged directly into the base model weights for standalone use. It is designed for image-text-to-text tasks and offers improved robustness or specific behavioral alignment due to the baked-in adversarial steering.

Loading preview...

Model Overview

groxaxo/gemma4-prometheus-merged is a 31 billion parameter model derived from google/gemma-4-31B-it. Its key differentiator is the integration of Prometheus adversarial-steering optimization, where the best steering vectors have been merged directly into the base model. This process results in a standalone model that does not require a separate Prometheus runtime, making it easier to deploy.

Key Capabilities and Features

  • Prometheus-Steered: Incorporates adversarial steering to potentially enhance model robustness or align its behavior according to the steering objectives.
  • Standalone Deployment: All steering adaptations are merged into the base weights, allowing for direct loading and inference without external frameworks.
  • Gemma-4 Architecture: Utilizes the gemma-4 architecture, which includes 60 transformer layers with alternating sliding-window and full attention mechanisms, and Grouped Query Attention (GQA).
  • Image-Text-to-Text Task: Intended for multimodal tasks involving image and text inputs to generate text outputs, though text-only inference is supported.
  • Context Length: Supports a base context length of 6,144 tokens, with an estimated potential for up to 226,000 tokens with flash-attn and FP8 KV cache.
  • Quantization Compatibility: Can be loaded in BF16 (requiring ~65 GiB VRAM) or 8-bit quantized mode (requiring ~48 GiB VRAM).

Use Cases and Considerations

This model is suitable for developers looking for a Gemma-4 based model with pre-applied adversarial steering for specific behavioral characteristics or enhanced robustness. It is particularly relevant for applications where the base Gemma-4 model's outputs need refinement through adversarial training. Users should be aware of the hardware requirements, especially for BF16 inference, and ensure enable_thinking=False is passed to apply_chat_template to prevent unintended chain-of-thought token generation.