plawanrath/mistral-7b-instruct-v0.3-bf16-mlx-cba

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:May 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

plawanrath/mistral-7b-instruct-v0.3-bf16-mlx-cba is a 7.2 billion parameter Mistral-family instruction-tuned causal language model, converted to BF16 precision for MLX (Apple Silicon) by Plawan Kumar Rath. This model serves as the uncompressed baseline artifact for research on quantization bias in LLMs, specifically investigating how compression affects emergent stereotypical behavior. It is optimized for direct loading and inference on MLX-compatible hardware, making it suitable for research into model quantization effects and efficient local deployment.

Loading preview...

Overview

This model, plawanrath/mistral-7b-instruct-v0.3-bf16-mlx-cba, is an MLX-format BF16 (uncompressed baseline) variant of the mistralai/Mistral-7B-Instruct-v0.3 model. Developed by Plawan Kumar Rath, it is one of 15 model artifacts used in the paper "Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels." This specific version serves as the uncompressed reference to study the impact of quantization on model behavior, particularly concerning bias.

Key Capabilities

  • MLX Native Format: Directly loadable and usable on Apple Silicon (MLX) without additional conversion steps.
  • BF16 Precision: Represents the uncompressed baseline for research into quantization effects.
  • Instruction-Tuned: Inherits the instruction-following capabilities of the base Mistral-7B-Instruct-v0.3 model.
  • Research Artifact: Provides the exact artifact used for inference results in a study on quantization bias, enabling reproducibility.

Good For

  • Research on Quantization Bias: Ideal for researchers investigating how model compression (quantization) affects emergent biases and stereotypical behavior in LLMs.
  • MLX-based Development: Developers working with MLX on Apple Silicon who need an instruction-tuned Mistral model in a native format.
  • Performance Baselines: Establishing uncompressed performance and behavior baselines before applying various quantization techniques.
  • Fairness-Sensitive Applications: Understanding the potential risks and changes in model behavior when deploying compressed instruction-tuned models in contexts where fairness is critical.

Finetunes
1 models