mrzenin/Decka-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

mrzenin/Decka-4B is a 4.5 billion parameter decensored version of Empero-AI's Qwen3.8-4B, built using the Heretic v1.4.0 tool. This model is a full-parameter distillation of a larger Qwen3.8 2.4T A95B teacher model, specifically optimized for reasoning, mathematics, and instruction following. It features a native 262,144-token context length and native function calling, designed to bring frontier-scale reasoning capabilities to consumer hardware.

Loading preview...

Overview

mrzenin/Decka-4B is a 4.5 billion parameter language model, a decensored variant of Empero-AI's Qwen3.8-4B, created using the Heretic v1.4.0 tool. It is a full-parameter distillation of an internal Qwen3.8 2.4T A95B teacher model, trained on approximately 45,000 curated teacher traces focusing on mathematics, general reasoning, and instruction following. The primary goal is to enable advanced reasoning behavior on consumer-grade hardware.

Key Capabilities

  • Distilled Chain-of-Thought: Utilizes <think> blocks learned directly from high-performing teacher model traces, rather than self-generated reasoning.
  • Efficient Size: At 4.5 billion parameters, it fits comfortably on consumer hardware (bf16 in ~8 GB, quantized builds on laptops).
  • Native Function Calling: Supports function calling as per Qwen3.5's specification without requiring additional fine-tuning.
  • Extended Context Window: Inherits a native 262,144-token context length from its Qwen3.5 base.
  • Improved MMLU Performance: Shows a significant improvement in MMLU (CoT, 57 subjects) scores compared to its Qwen3.5-4B base, with a +0.199 increase in flexible-extract accuracy.

Good For

  • Reasoning Tasks: Excels in tasks requiring mathematical problem-solving and general reasoning due to its distilled chain-of-thought approach.
  • Instruction Following: Optimized for complex instruction adherence, making it suitable for various NLP applications.
  • Resource-Constrained Environments: Its efficient parameter count allows deployment on consumer GPUs and laptops.
  • Applications Requiring Decensored Outputs: Provides a less restricted output compared to its original counterpart, suitable for specific research or creative applications.

Finetunes
2 models