OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

OliviaRossi/MiMo-Ornith-9B-AGSI-Abliterated-HQ is a 9.7 billion parameter Qwen3_5-based model, developed by OliviaRossi, engineered for high-precision reasoning and agentic capabilities. It uniquely combines chain-of-thought reasoning from MiMo-V2.6 with reinforcement-learned terminal execution from Ornith-1.5, using Adaptive Geodesic Spectral Interpolation (AGSI) for merging. The model features Bayesian Representation Abliteration (Abliterix) to remove refusal vectors while preserving reasoning, making it suitable for complex algorithmic synthesis, coding, and multi-step problem-solving without censorship.

Loading preview...

MiMo-Ornith-9B-AGSI-Abliterated-HQ: High-Precision Reasoning & Agentic Hybrid

This 9.7 billion parameter model, developed by OliviaRossi, is a sophisticated merge designed for advanced reasoning and agentic tasks. It integrates the strengths of two base models:

  • Xiaomi MiMo-V2.6-Distill-Qwen-9B: Contributes chain-of-thought reasoning, complex mathematical problem-solving, and deep algorithmic synthesis.
  • DeepReinforce Ornith-1.5-9B: Provides reinforcement-learned terminal execution, autonomous bash commands, multi-step tool calling, and self-correcting agentic loops.

Key Differentiators & Features

  • Adaptive Geodesic Spectral Interpolation (AGSI): A custom parameter-space merge technique that preserves individual neuron manifolds and activation energy, preventing signal degradation common in standard merging methods. It intelligently routes contributions across layers, with MiMo-dominance in early layers for syntactic stability and Ornith-dominance in later layers for agentic tool use.
  • Bayesian Representation Abliteration (Abliterix): Utilizes a multi-objective Optuna TPE search to orthogonally project out refusal vectors, ensuring the model is uncensored while strictly maintaining reasoning and syntax precision (KL-divergence constraint $\le 0.007$). This avoids the "lobotomization" effect of simpler refusal removal methods.
  • Native GGUF Compatibility: Addresses and fixes upstream Qwen 3.5 MTP configuration bugs, ensuring 100% plug-and-play GGUF conversion for llama.cpp and vLLM without requiring special flags.
  • Configurable Reasoning Effort: The chat_template allows users to dynamically steer the model's reasoning depth from none (instant generation) to max (exhaustive hypothesis exploration), optimizing for latency or rigor based on the task.
  • Prefix KV Cache Parity: Supports preserve_thinking to maintain historical <think> traces, preventing multi-turn context drift and ensuring 100% prefix KV cache hit rates for consistent long-context interactions.

Ideal Use Cases

  • Complex Algorithmic Synthesis: For generating intricate code and algorithms.
  • Security Research & Penetration Testing: Due to its uncensored nature and reasoning capabilities.
  • Multi-step Problem Solving: Excels in tasks requiring sequential thought and tool invocation.
  • Agentic Workflows: For autonomous execution of commands and self-correcting loops.

Finetunes
2 models