Mantisec/Qwen3.8-27B-OBLITERATED-FP16

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026Architecture:Transformer Featherless Exclusive Cold

Mantisec/Qwen3.8-27B-OBLITERATED-FP16 is a 27 billion parameter language model, converted to FP16 from OBLITERATUS/Qwen3.8-27B-OBLITERATED by Mantisec using bfsquish v0.1.0. This model is specifically optimized for efficient inference and fine-tuning on NVIDIA V100 (Volta) GPUs, which lack native BF16 Tensor Core support. It maintains high numerical quality with 100% token agreement and minimal logit drift compared to its source, making it suitable for applications requiring FP16 precision on target hardware.

Loading preview...

Overview

Mantisec/Qwen3.8-27B-OBLITERATED-FP16 is a 27 billion parameter model derived from OBLITERATUS/Qwen3.8-27B-OBLITERATED. This version has been precisely converted to FP16 (16-bit floating point) using the bfsquish tool by Mantisec, specifically targeting NVIDIA V100 (Volta) GPUs. The conversion employs a range_checked strategy, ensuring that weights are accurately represented in FP16 without clipping values outside its finite range, thus preserving the source checkpoint's integrity.

Key Capabilities

  • Optimized for NVIDIA V100 GPUs: Designed for efficient inference and fine-tuning on Volta architecture, which benefits from FP16 precision.
  • High Numerical Fidelity: Validation metrics show a 100% token agreement rate and a minimum cosine similarity of 0.999983 compared to the source BF16 model, indicating excellent preservation of numerical quality despite the format change.
  • Robust Conversion: The range_checked strategy prevents data loss by rejecting values that cannot be accurately represented in FP16, ensuring reliable performance.

Good For

  • Developers and researchers working with NVIDIA V100 GPUs who require optimized FP16 models for faster processing and reduced memory footprint.
  • Fine-tuning existing models on hardware that performs better with FP16 data types.
  • Applications where maintaining high numerical accuracy during precision conversion is critical.