anjohn0077/NEXS-qwen3-32b-multislerp

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026Architecture:Transformer Featherless Exclusive Cold

The anjohn0077/NEXS-qwen3-32b-multislerp is a 32 billion parameter language model based on the Qwen3ForCausalLM architecture, created by anjohn0077 using a multi-SLERP merge technique. This model integrates specialized domain expertise in instruction-following, medical knowledge, and Russian language capabilities. It is designed to offer a versatile solution for tasks requiring proficiency across these distinct domains.

Loading preview...

NEXS Qwen3-32B Multi-SLERP Merge Overview

This model, developed by anjohn0077, is a 32 billion parameter large language model built upon the Qwen3ForCausalLM architecture. It stands out due to its unique construction via a multi-SLERP merge method, which combines the strengths of several specialized "domain expert" models. This technique involves barycentric spherical interpolation in task-vector space, projecting model deltas from a shared base into tangent space, interpolating, and then projecting back.

Key Capabilities & Merged Expertise

The NEXS Qwen3-32B Multi-SLERP model integrates expertise from three distinct sources, each contributing to its specialized capabilities:

  • Instruction-Following: Enhanced ability to understand and execute complex instructions, derived from qihoo360/Light-IF-32B.
  • Medical Domain Knowledge: Proficient in medical-related queries and tasks, incorporating knowledge from OpenMedZoo/MedGo.
  • Russian Language Proficiency: Strong capabilities in processing and generating Russian text, sourced from t-tech/T-pro-it-2.0.

Technical Details

The merge process utilized mergekit with multislerp method, using Qwen/Qwen3-32B as the base model. All source models were given equal weight during the spherical averaging of their task-vector deltas. The tokenizer was reconciled to the base model's tokenizer to handle minor vocabulary differences among the variants.

Good for

  • Applications requiring strong instruction-following across various tasks.
  • Use cases in the medical field that demand specialized language understanding.
  • Projects needing robust Russian language processing capabilities.

Finetunes
1 models