JuanCDEV/rovalthia

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

JuanCDEV/rovalthia, branded as ROLLY-AGENTIC-CYBER-12-9T, is a 4.66 billion parameter model built on the Qwen3.5 architecture, developed by ROVALTHIA LABORATORY. This model is designed with a strong focus on agentic intelligence and defensive cybersecurity, featuring a sophisticated memory mesh, skill system, and knowledge factory. It is intended for use in complex reasoning, tool use, and cybersecurity applications, with an ambitious target context window of 1 million tokens.

Loading preview...

ROLLY-AGENTIC-CYBER-12-9T: Agentic Intelligence for Cybersecurity

ROLLY-AGENTIC-CYBER-12-9T, developed by ROVALTHIA LABORATORY, is a 4.66 billion parameter model based on the Qwen3.5 architecture. While its physical parameter count is 4.66B, it is designed to scale to an 11.49B parameter target architecture. The model features an ambitious 1 million token context window, enabled by YaRN RoPE.

Key Capabilities & Features

  • Agentic Architecture: Incorporates a sophisticated "Memory Mesh V3" with 10 isolated scopes (e.g., WORKING, CONVERSATION, EPISODIC, SKILL) and a "Skill System V2" with 17 skill families, governed by an EconomyRoutingSkill for efficient compute. This system aims for minimal agent hops and optimized internal token reduction.
  • Cybersecurity Focus: Includes a "Knowledge Factory" with ingested research papers on Zero Trust and CTI threat hunting, and a dedicated training curriculum domain for Defensive Cyber.
  • Structured Training Curriculum: Designed with 17 specialized domains, including Instruction, Mathematics, Reasoning, Coding, Tool Use, Agent Routing, and Defensive Cyber. The model is currently in a STAGED_READY_FOR_GPU_CLUSTER state, awaiting full training on enterprise GPU nodes.
  • Forensic Verification: Detailed tensor-level audit provided, including SHA-256 hashes for model integrity.

Good For

  • Autonomous Agent Development: Its advanced memory and skill systems are tailored for building complex, multi-agent applications.
  • Defensive Cybersecurity: Specialized knowledge and training domains make it suitable for tasks like threat modeling, security analysis, and incident response.
  • Complex Reasoning & Tool Use: Designed to excel in tasks requiring structured output, tool integration, and agent planning, with internal benchmarks showing 100% on Instruction Following and Defensive Cyber diagnostics.

Finetunes
1 models