Level 0: The Model

The Model

Direct access to TasteBrain weights and the low-level multi-vector interface for advanced developers and researchers.

Overview

Level 0 provides the foundational model layer of TasteBrain. This level is designed for:

  • Advanced developers and researchers who download models from Hugging Face
  • Teams running local inference or operating directly on raw vectors
  • Scientific exploration of high-dimensional "taste vectors"

Core Utility

Craft custom late-binding queries and experiment with high-dimensional "taste vectors" for scientific or proprietary style analysis.

Model Architecture

TasteBrain is built on Qwen3, a multi-modal foundation model, trained with contrastive learning on our proprietary dataset of aesthetic attributes.

Key Features

  • Multi-Modal Embeddings - Unified embedding space for images, text, and products
  • Aesthetic Manifold - High-dimensional space encoding "DNA of taste"
  • Late-Binding Queries - Flexible multi-vector matching at inference time
  • Cross-Domain Search - Map audio to products, food preferences to products, etc.

Deployment

Model cards include structured metadata for:

  • Task identification - e.g., aesthetic-understanding, style-transfer
  • Base-model tracking - Version control and lineage
  • Performance benchmarks - Accuracy metrics on curated test sets

Current Usage

Level 0 is currently used exclusively by internal teams at Bestomer for:

  • Research and development of new aesthetic understanding capabilities
  • Direct access to vectors for operating on manifolds
  • Custom query optimization and experimentation
  • Training pipeline development

⚠️ Limited Public Access

Level 0 access is restricted to research partners and advanced use cases. For most developers, Level 1: SVK Primitives provides the right abstraction layer.

Next Steps