Blackfrost-AI/PINQWEN-3.6-35B-CLEAN-BF16

PINQWEN-3.6-35B-CLEAN-BF16 Overview

PINQWEN-3.6-35B-CLEAN-BF16 is a 35.1 billion parameter Mixture-of-Experts (MoE) model developed by Blackfrost-AI, based on Alibaba's Qwen/Qwen3.6-35B-A3B. This model utilizes a qwen3_5_moe architecture with 256 experts (8 routed + 1 shared per token) and approximately 3 billion active parameters per token, featuring a hybrid Gated-DeltaNet and gated attention. It is the full 16-bit merged weights of the supervised fine-tuned (SFT) version, designated as 'CLEAN' for its preserved safety alignment.

Key Capabilities & Training

• General-purpose assistant: Designed for broad utility in text-based tasks.
• Optimized for reasoning: Fine-tuned to excel in complex logical deduction and problem-solving.
• Agentic tool-use: Specifically trained for ReAct-style tool integration and execution.
• Proprietary fine-tuning: SFT was conducted on Blackfrost-AI's proprietary "The Void (v4)" corpus, comprising over 5,000 high-quality multi-turn examples blending reasoning, broad knowledge, and agentic trajectories.
• Robust training: Utilized bf16 LoRA with Unsloth, applying adapters to attention projections and MoE expert projections, resulting in 1.86 billion trainable parameters (5.04% of the model).

Intended Uses & Limitations

• Intended Uses: Primarily serves as a general-purpose text assistant for reasoning and agentic/tool-use tasks.
• Text-focused SFT: While the base model is vision-language, vision capabilities were not specifically tuned or evaluated in this fine-tuning process.
• No public benchmarks: Internal evaluations are ongoing, and no public benchmark scores are currently reported.
• Potential for hallucination: Like all large language models, it may inherit base-model limitations and can hallucinate; verification of important facts is recommended.