osk-arr00/BigBang-Aquila-35B-Merged
The osk-arr00/BigBang-Aquila-35B-Merged is a 35.1 billion parameter multimodal language model, based on Qwen/Qwen3.6-35B-A3B, created using a DARE-TIES merge method. It combines two specialized models: one for formal reasoning, math, code, and research, and another for web agency tasks like deep search, scraping, and tool-calling. This merge aims to retain both capabilities, making it suitable for complex tasks requiring both analytical and web interaction skills.
Loading preview...
Model Overview
The osk-arr00/BigBang-Aquila-35B-Merged is a 35.1 billion parameter multimodal language model derived from Qwen/Qwen3.6-35B-A3B. It was created using the DARE-TIES (Drop And REscale + Trimming, Electing Sign & Disjoint Merge) method, combining two distinct specialist models.
Key Capabilities
- Formal Reasoning & Technical Tasks: Inherits strong capabilities in math, code, and research from
endless-frontier/BigBang-v1. - Web Agency & Tool-Calling: Excels at deep search, web scraping, and tool-calling, derived from
XYZAILab/XYZ-Aquila-mini. - Multimodal: Supports image-text-to-text processing, with the core LLM residing under
model.language_model.*. - Mergeability: The DARE-TIES method ensured that the capabilities of both merged models were largely retained, indicated by a low cosine similarity (0.057) between their parameter tensors.
Evaluation & Performance
Agentic evaluations across 9 hard cases showed a 6/9 PASS rate, with strong performance in BigBang-specific domains (math/code) and reasonable performance in web agency and intersectional tasks.
Good For
- Applications requiring a blend of advanced reasoning, mathematical problem-solving, and code generation.
- Tasks involving web interaction, such as automated data extraction, intelligent search, and complex tool orchestration.
- Multimodal scenarios where both text and image inputs are processed for text-based outputs.