nightmedia/Qwen3.6-35B-A3B-Qwable-Holo3-Qwopus
nightmedia/Qwen3.6-35B-A3B-Qwable-Holo3-Qwopus is a 35.1 billion parameter language model created by nightmedia, merged using NuSLERP from several Qwen-based models including Qwable-v1 and Qwen3.6-35B-A3B-MTP-Holo3-Qwopus. This model features a 32K context length and incorporates a unique 'thinking toggle' mechanism via control tokens for explicit control over reasoning depth, making it suitable for applications requiring both quick answers and deep analytical processing.
Loading preview...
Model Overview
nightmedia/Qwen3.6-35B-A3B-Qwable-Holo3-Qwopus is a 35.1 billion parameter model, a NuSLERP merge of several Qwen-based architectures, including llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved, samuelcardillo/Qwopus-MoE-35B-A3B, Hcompany/Holo3-35B-A3B, and lordx64/Qwable-v1. It is designed with a 32,768 token context length.
Key Capabilities & Features
- NuSLERP Merge: Combines the strengths of multiple Qwen-based models, aiming for enhanced performance across various tasks.
- "Thinking Toggle" Mechanism: Integrates a unique feature allowing users to explicitly control the model's reasoning process using
<|think_on|>and<|think_off|>control tokens. This enables switching between fast, direct answers and deeper, more elaborate reasoning without the model seeing the control tokens in context. - Preserve Thinking Flag: Includes
<|think_forget|>and<|think_remember|>tags to manage thepreserve_thinkingflag, offering further control over the model's internal state during reasoning. - Performance Metrics: Benchmarks provided for various tasks (arc, boolq, hswag, obkqa, piqa, wino) across different quantization levels (bf16, mxfp8, qx86-hi, qx64-hi, mxfp4), showing competitive results.
When to Use This Model
This model is particularly well-suited for applications where dynamic control over the model's reasoning depth is beneficial. Developers can leverage the "thinking toggle" for:
- Code Generation & Complex Problem Solving: Use
<|think_on|>for detailed, step-by-step reasoning. - Quick Q&A & Chatbots: Employ
<|think_off|>for concise, immediate responses. - Adaptive AI Systems: Build systems that can adjust their response style based on user input or task requirements, switching between analytical and direct modes.