Merlin-Research/Qwen3.5-4B-Safety-Thinking

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Loading

Merlin-Research/Qwen3.5-4B-Safety-Thinking is a 4 billion parameter language model developed by Merlin Research, based on the Qwen/Qwen3.5-4B architecture. It is specifically optimized for structured reasoning, strict instruction adherence, and safety-aligned behavior, with a potential context length of 1 million tokens. This model excels in applications requiring robust, predictable, and safe autonomous behavior, particularly in assistant and agent workflows.

Loading preview...

Model Overview

Merlin-Research/Qwen3.5-4B-Safety-Thinking is a 4 billion parameter model developed by Merlin Research, built upon the Qwen/Qwen3.5-4B base. It has been rigorously optimized through LoRA-based Supervised Fine-Tuning (SFT) to enhance safety reasoning, controllability, and response consistency.

Key Capabilities

  • Structured Reasoning Quality: Enhanced ability to process complex problems step-by-step.
  • Instruction Adherence: Superior capability to follow strict guidelines and constraints.
  • Safety-Aligned Behavior: Designed for safe operation in practical assistant and autonomous agent workflows.
  • Robustness: Increased resistance against misalignment patterns and adversarial inputs.
  • Native Reasoning Format: Supports and normalizes the <think>...</think> format to explicitly separate reasoning from the final output.

Training and Data

The model leverages a rigorous post-training stack, combining supervised reasoning tuning with alignment-oriented optimization. It was trained on Merlin Research's private datasets, focusing on reasoning reliability, instruction-following robustness, safety behavior refinement, and misalignment reduction, utilizing Anthropic's Bloom&Petri framework for behavioral alignment.

Intended Use Cases

This model is particularly well-suited for:

  • Building safety-oriented reasoning assistants and chatbots.
  • Tasks requiring strict, constrained instruction-following.
  • Experimentation in AI alignment, safety research, and robustness testing.
  • Agentic workflows demanding predictable and safe autonomous behavior.