oktayd/Q36-35B-A3B-Opus4.7-Ablit-Heretic-OBLITERATUS-Hermes-MTP-Vision-FT

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

oktayd/Q36-35B-A3B-Opus4.7-Ablit-Heretic-OBLITERATUS-Hermes-MTP-Vision-FT is a 35.1 billion parameter Qwen3.6-35B-A3B model with 3 billion active parameters, developed by oktayd. This full merged BF16 model preserves vision and Multi-Tool Protocol (MTP) capabilities, making it suitable for image-text-to-text and text generation tasks. It functions as a master checkpoint and server model, optimized for FreeToken use, and demonstrates strong performance in instruction following and reasoning benchmarks.

Loading preview...

Model Overview

oktayd/Q36-35B-A3B-Opus4.7-Ablit-Heretic-OBLITERATUS-Hermes-MTP-Vision-FT is a 35.1 billion parameter model based on the Qwen3.6-35B-A3B family, featuring 3 billion active parameters. This BF16 model is designed for Image-Text-to-Text and text generation, with preserved Vision and Multi-Tool Protocol (MTP) capabilities. It serves as a master checkpoint, optimized for FreeToken environments.

Key Capabilities & Features

  • Hybrid Architecture: Utilizes a qwen3_5_moe architecture with 35B total parameters and 3B activated, balancing performance and efficiency.
  • Multimodal Support: Retains full Vision and MTP functionalities, enabling complex multimodal interactions.
  • Extensive Lineage: Built upon a sophisticated lineage including Qwen3.6, reasoning-distilled derivatives, abliterated stages, custom fused-MoE-aware Heretic, OBLITERATUS Nuclear, and Hermes Function Calling SFT.
  • Robust Training: Trained with 24,399 SFT examples, ensuring high integrity with zero parser errors and protected Vision/MTP tensors.
  • Benchmark Performance: Demonstrates strong results in instruction following (IFEval strict instruction: 94.4%), output integrity, and HumanEval+ coding tasks. Shows a completion/reasoning-robustness advantage in MATH-500.

Ideal Use Cases

  • Multimodal Applications: Excellent for tasks requiring both image and text understanding, leveraging its preserved Vision and MTP capabilities.
  • Instruction Following: Suited for applications where precise adherence to instructions is critical, as indicated by its high IFEval scores.
  • Agentic Workflows: Benefits from its Hermes Function Calling and Agent/coding/terminal/file/repo/multi-tool SFT, making it suitable for complex agentic tasks.
  • Reasoning & Problem Solving: Its reasoning-distilled lineage and performance in MATH-500 suggest strong capabilities for analytical and problem-solving scenarios.

Finetunes
1 models