sovthpaw/omnistep-new

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

sovthpaw/omnistep-new is an 8.2 billion parameter omnimodal model developed by SouthpawIN, featuring a Qwen3-8B text backbone, Cosmos multimodal heads for image and video understanding, and ACE-Step music modules for music generation from lyrics. It supports a 40K native context length, extendable to 256K+, and integrates tool/function calling. This model is designed as a versatile base for multimodal applications, particularly excelling in combined chat, visual comprehension, and music creation tasks.

Loading preview...

OmniStep-new: An Omnimodal Music Generator + Chat Assistant

OmniStep-new, developed by SouthpawIN, is an 8.2 billion parameter omnimodal model designed to be a versatile base for advanced AI applications. It integrates multiple capabilities into a single bundled model, making it suitable for a wide range of multimodal tasks.

Key Capabilities

  • Chat / Instruction Following: Powered by a Qwen3-8B text backbone, it handles general conversational and instruction-based tasks.
  • Image + Video Understanding: Utilizes Cosmos multimodal heads for processing and understanding visual inputs, integrated via specific chat template tokens.
  • Music Generation: Features ACE-Step v1.5 turbo DiT + 1.7B LM + VAE modules, enabling the generation of music from lyrics.
  • Tool / Function Calling: Supports OpenAI-style tool schemas through <tool_call> tags in its chat template.
  • Long Context: Offers a 40K native context length, which can be extended to over 256K using YaRN.

Architecture and Development

This model was built using Darwin Family weight-space merging (MRI-Trust Fusion), combining nvidia/Cosmos3-Nano and Qwen/Qwen3-8B. It represents the foundational 'gen 0' model in the OmniSenter family, intended for SFT warm-start or direct inference. Future iterations include an agentic SFT variant and a sparse-upcycled MoE model.

Good For

  • Developers looking for a unified model capable of text chat, image/video understanding, and music generation.
  • Applications requiring long context processing and tool-use capabilities.
  • As a base model for further fine-tuning (SFT) for agentic behaviors.

Finetunes
1 models