VertexAGI/amethyst-1-mini

VISIONPricing:Input $0.2 / Output $0.4Concurrent Unit Cost:1Model Size:4.3BQuant:BF16Context Size:32kPublished:Aug 7, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

VertexAIco/amethyst-1-mini is a 4 billion parameter general-purpose chat and instruction-following model, fine-tuned from Google's Gemma 3 4B IT. Developed by an independent research project, it was created using LoRA on a small, high-quality distilled instruction dataset. This model is designed for experimentation, research into small-scale distillation pipelines, and hobbyist deployment, focusing on validating an end-to-end distillation and fine-tuning pipeline on consumer hardware.

Loading preview...

Amethyst 1 Mini: Distilled Instruction-Following Model

Amethyst 1 Mini is a 4 billion parameter, general-purpose chat and instruction-following model, developed as an independent research project. It is fine-tuned from Google's Gemma 3 4B IT using LoRA, representing the first model in the Amethyst family. The primary goal of this release is to validate an end-to-end distillation, fine-tuning, and evaluation pipeline on consumer hardware.

Key Capabilities & Training:

  • Base Architecture: Utilizes the Gemma 3, 4B parameter dense, decoder-only transformer architecture.
  • Distilled Data: Fine-tuned on 1,122 high-quality instruction/response pairs, synthetically generated via knowledge distillation from nvidia/nemotron-3-super-120b-a12b.
  • Broad Coverage: The training dataset spans diverse general-chat categories, including reasoning, code generation, creative writing, and translation, aiming for well-rounded conversational ability.
  • Fine-tuning Method: Employs LoRA (rank 8, scale 20.0) on Apple Silicon using the MLX framework, with the adapter fused into the base weights and dequantized to fp16 for direct transformers compatibility.

Intended Use & Limitations:

  • Good for: Experimentation, research into small-scale distillation pipelines, and hobbyist deployment. It serves as a lightweight, general-purpose conversational assistant.
  • Limitations: Due to its small, synthetic training dataset, behavior can be inconsistent outside trained categories. It inherits the limitations and knowledge cutoff of its Gemma 3 4B IT base model and is not intended for high-stakes or production use.