DuoNeural/Archon-14B
Archon-14B by DuoNeural is a 14.7 billion parameter language model based on Qwen3-14B, featuring a 32K context length. It is specifically modified to remove refusal behaviors while retaining its built-in chain-of-thought reasoning via blocks. This model excels at code, math, and multilingual tasks, offering a powerful, unrestricted thinking model suitable for users with 16GB VRAM.
Loading preview...
DuoNeural/Archon-14B: Unrestricted Reasoning Model
Archon-14B is a 14.7 billion parameter model developed by DuoNeural, built upon Alibaba's Qwen3-14B base. It retains the Qwen3 series' strong capabilities in code, mathematics, and multilingual tasks, along with a built-in chain-of-thought reasoning mechanism using <think> blocks. The model's primary differentiator is its abliteration of refusal behaviors through a Single-pass BF16 SVD process on an NVIDIA A6000, which removed safety conditioning while preserving its reasoning abilities.
Key Capabilities & Features
- Unrestricted Reasoning: Modified to remove refusal directions, allowing for unconstrained thought processes.
- Built-in Chain-of-Thought: Utilizes
<think>blocks for enhanced reasoning, which can be optionally disabled for faster responses. - Strong Performance: Inherits Qwen3-14B's proficiency in code generation, mathematical problem-solving, and multilingual understanding.
- Optimized for Consumer Hardware: Runs in 4-bit NF4 quantization with approximately 9GB VRAM, making it suitable for GPUs like the RTX 3090/4090.
- Apache 2.0 License: Offers flexible usage for developers.
Ideal Use Cases
- Complex Problem Solving: Leveraging its chain-of-thought and unrestricted nature for intricate tasks.
- Code Generation & Analysis: Benefiting from its strong coding capabilities.
- Multilingual Applications: For tasks requiring proficiency across various languages.
- Creative & Unfiltered Content Generation: Where traditional safety filters might hinder output.
Archon-14B is positioned as a "sweet spot" in the Archon series, balancing capacity and reasoning with accessibility for consumer-grade GPUs.