Qwen3-4B-Qwen3.6-plus-Reasoning-Slerp/README.md

base_model, library_name, tags, license, pipeline_tag, language, datasets

base_model	library_name	tags	license	pipeline_tag	language	datasets
khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Distilled	transformers	mergekit merge	apache-2.0	text-generation	en	khazarai/kimi-2.5-high-reasoning-250x khazarai/qwen3.6-plus-high-reasoning-500x

khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Slerp

Note: The sharp drop in "Creative Writing" is an expected and accepted trade-off to maximize extreme logical reasoning and coding precision.

This model is a highly experimental and optimized reasoning model created through a surgical SLERP merge of two powerful 4B reasoning models. The goal of this merge was to combine the deep analytical capabilities of Kimi with the mathematical and structural precision of Qwen, while mitigating the catastrophic forgetting commonly seen in SFT model merges. After multiple iterations and layer-by-layer tensor analysis, we achieved a "1+1=3 Synergy Effect" in Logical Inference and Planning, outperforming both base models and the official Qwen Thinking model.

The "Golden Path" (V5) Strategy

Standard SLERP merges often destroy RAG capabilities and syntax adherence. To solve this, this model utilizes a custom merge configuration:

1. RAG/Vocabulary Fix: embed_tokens and lm_head are strictly pinned to 1.0 (Qwen). The model reads and speaks purely using Qwen's vocabulary, completely eliminating the RAG degradation problem.
2. Gradient Attention: The intermediate attention and MLP layers follow a smooth gradient [0, 0.1, 0.2, 0.3, 0.5, 0.7, 0.8, 0.9, 1] to prevent weight interference in deep reasoning steps.

Benchmark Performance (Multi-Domain Reasoning)

Model	Score
khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Slerp	77.18
khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled	76.09
khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Distilled	75.64
Qwen/Qwen3-4B-Thinking-2507	73.73

• Benchmark: khazarai/Multi-Domain-Reasoning-Benchmark
• Total Questions: 100

💡 Intended Use Cases

• Ideal for: Complex logical deductions, Python code debugging, mathematical problem-solving, and strict RAG (Retrieval-Augmented Generation) pipelines.
• Not recommended for: Creative writing, poetry, or highly imaginative storytelling.

Models Merged

The following models were included in the merge:

• khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled
• khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Distilled

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled
  - model: khazarai/Qwen3-4B-Qwen3.6-plus-Reasoning-Distilled
merge_method: slerp
base_model: khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled
parameters:
  t:
    - filter: embed_tokens
      value: 1
      
    - filter: lm_head
      value: 1

    - value: 1
    
    - filter: self
      value: [0, 0.1, 0.2, 0.3, 0.5, 0.7, 0.8, 0.9, 1]
      
dtype: bfloat16