AMAImedia/NOESIS-Gemma4-12B-it-Qat-Q4_0-Unquantized-BF16

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 16, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

AMAImedia/NOESIS-Gemma4-12B-it-Qat-Q4_0-Unquantized-BF16 is a 12 billion parameter instruction-tuned Gemma 4 model, developed by AMAImedia as part of their NOESIS Professional Multilingual Dubbing Automation Platform. This model serves as a wired Scenema text-encoder, specifically integrated via a non-linear G4→G3 hidden-state aligner to bridge compatibility with the Scenema-DiT audio generation system. It excels in multilingual dubbing applications, achieving a mean per-language validation cosine similarity of 0.908 across 12 of 13 languages, making it suitable for high-quality audio dubbing workflows.

Loading preview...

Model Overview

This model, NOESIS-Gemma4-12B-it-Qat-Q4_0-Unquantized-BF16, is a 12 billion parameter instruction-tuned Gemma 4 model developed by AMAImedia as a core component of their NOESIS Professional Multilingual Dubbing Automation Platform. It functions as a wired Scenema text-encoder, specifically designed to integrate Gemma 4's output with the Scenema-DiT audio generation system.

Key Capabilities & Features

  • Gemma 4 Architecture: Utilizes Google's next-generation Gemma 4 (12B instruction-tuned) architecture.
  • Hidden-State Alignment: Features a unique, non-linear G4→G3 hidden-state aligner (scenema_aligner_v1/) that maps Gemma 4's output hidden states onto the Gemma 3 distribution expected by Scenema-DiT. This aligner is applied at inference and cannot be weight-merged into the base model.
  • Multilingual Performance: Achieves a mean per-language validation cosine similarity of 0.908, with 12 of 13 languages (e.g., es, pt, de, it, ru, fr, en, sw, hi, zh) performing around 0.92-0.93. Korean (ko) is noted as a weak outlier at 0.728.
  • BF16 Precision: The model uses verbatim upstream BF16 weights.
  • Context Length: Supports a context length of 32768 tokens.

Good For

  • Professional Multilingual Dubbing: Ideal for integration into the NOESIS platform for high-quality audio dubbing automation.
  • Research & Development: Serves as a substrate for cloud-side adapter experiments and further development within the Scenema family of models.
  • Applications Requiring Gemma 4 Integration: Useful for scenarios where Gemma 4's capabilities need to be adapted for compatibility with systems previously trained on Gemma 3 hidden-state distributions.

Finetunes
1 models