cryptonaut/Qwen3.5-noholds-35b-a3b

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

cryptonaut/Qwen3.5-noholds-35b-a3b is a 35.1 billion parameter causal language model, a decensored version of Qwen-AgentWorld-35B-A3B, developed by cryptonaut using Heretic v1.4.0. This model functions as a native language world model, specifically trained to simulate agentic environments across seven interaction domains. It excels at predicting next environment states given agent actions and interaction history, making it suitable for agentic environment simulation and research.

Loading preview...

Qwen3.5-noholds-35b-a3b: Decensored Language World Model

This model is a decensored variant of the Qwen-AgentWorld-35B-A3B, created using Heretic v1.4.0. It is a native language world model designed to simulate agentic environments by predicting the next environment state based on an agent's actions and interaction history. Unlike general-purpose LLMs, environment modeling is its core training objective from the initial stages.

Key Capabilities

  • Seven Unified Domains: Covers MCP (tool calling), Search, Terminal, SWE (software engineering), Android, Web, and OS, spanning both text and GUI interaction environments.
  • Agentic Environment Simulation: Trained through a three-stage pipeline (CPT, SFT, RL) to activate next-state-prediction reasoning and sharpen simulation fidelity.
  • Reproducible Decensoring: The model's decensoring process is reproducible, with specific abliteration parameters detailed.
  • Reduced Refusals: Demonstrates a lower refusal rate (41/100) compared to its original counterpart (57/100).
  • Extended Context Length: Features a substantial context length of 262,144 tokens, recommended to be at least 128K for optimal multi-turn environment simulation.

Good For

  • Agentic Research and Development: Ideal for researchers and developers building and testing AI agents in simulated environments.
  • Environment Modeling: Simulating complex interactions across diverse domains like software engineering, web browsing, and operating systems.
  • Applications Requiring Decensored Outputs: Suitable for use cases where less restrictive content generation is desired, while maintaining world model capabilities.