newmindai/Llama-3.1-8B-Instruct-w16a8-rw

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Nov 6, 2025License:llama3.1Architecture:Transformer Featherless Exclusive Cold

The newmindai/Llama-3.1-8B-Instruct-w16a8-rw is an 8 billion parameter, instruction-tuned variant of Meta's Llama-3.1-8B-Instruct model, specifically adapted for Turkish legal reasoning. It was trained using a Float8 rowwise quantization recipe, achieving a 19.87% speedup over BF16 baselines on H100 GPUs while maintaining stable convergence. This model excels in legal question-answering within Turkish law, leveraging its specialized training on the newmindai/EuroHPC-Legal dataset.

Loading preview...

Model Overview

This model, newmindai/Llama-3.1-8B-Instruct-w16a8-rw, is an 8 billion parameter instruction-tuned variant of Meta's Llama-3.1-8B-Instruct. It has been specifically adapted for Turkish legal reasoning, making it a specialized tool for legal domain applications.

Key Differentiators & Technical Details

  • Domain Adaptation: Fine-tuned on the newmindai/EuroHPC-Legal dataset, which consists of multi-domain Q/A in Turkish law, to enhance its legal reasoning capabilities.
  • Quantization for Efficiency: Utilizes a Float8 rowwise quantization recipe (w16a8-rw) during training. This method keeps model weights in BF16 but dynamically quantizes inputs, weights, and gradient outputs to FP8-E4M3 with a rowwise scaling configuration.
  • Training Speedup: Achieved approximately 19.87% speedup over BF16 baselines on H100 GPUs, demonstrating improved training efficiency without compromising convergence stability.
  • Base Model Architecture: Built upon the Llama 3.1 architecture with a notable context length of 32,768 tokens (though trained with a sequence length of 4,096 for this specific adaptation).

Intended Use Cases

  • Turkish Legal Reasoning: Optimized for question-answering and reasoning tasks within various subfields of Turkish law.
  • Research in Quantization: Suitable for researchers exploring the impact of fine-grained FP8 scaling on training efficiency and downstream performance in domain-adapted LLMs.

Limitations

  • Primarily for research and development; not a substitute for professional legal advice.
  • Users must ensure compliance with data protection and sector-specific regulations.
  • Potential for biases inherited from the domain-specific training data and model outputs.

Finetunes
6 models