schonsense/70B_llama33_stock_unslop

TEXT GENERATIONConcurrency Cost:4Model Size:70BQuant:FP8Ctx Length:32kPublished:Feb 28, 2026Architecture:Transformer Cold

schonsense/70B_llama33_stock_unslop is a 70 billion parameter language model based on the Llama 3.1 architecture, created by schonsense using a Model Stock merge. This model combines Llama 3.3 Instruct variants, including an 'abliterated' version and a 'slop token suppressed' version, aiming to achieve a balanced performance. It is designed to integrate the strengths of its constituent models, offering a versatile base for various generative AI tasks.

Loading preview...

schonsense/70B_llama33_stock_unslop Overview

This 70 billion parameter model, developed by schonsense, is an experimental merge created using the Model Stock method. It is built upon the meta-llama/Llama-3.1-70B base model and integrates several Llama 3.3 Instruct variants to achieve a balanced and robust performance profile.

Key Capabilities & Merge Details

The model aims to combine the best attributes of its constituent models, which include:

  • sam-paech/Llama-3.3-70B-Instruct-ftpo_1k
  • meta-llama/Llama-3.3-70B-Instruct
  • schonsense/llama33_inst_multivector_derestriction

The merge process utilized mergekit with specific parameters, including int8_mask: true and dtype: float32 (output bfloat16), to optimize the integration of these diverse Llama 3.3 Instruct versions. The goal is to mitigate potential downsides of individual models while enhancing overall utility.

Current Status

This model is currently in its testing phase, indicating ongoing evaluation and refinement. Developers should consider its experimental nature when deploying for critical applications.