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