OliviaRossi/MiMo-Ornith-9B-AGSI

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

OliviaRossi/MiMo-Ornith-9B-AGSI is a 9 billion parameter model based on the Qwen 9B architecture, created by OliviaRossi. It is a non-linear synthesis of two specialized models, MiMo-V2.6-Distill-Qwen-9B and Ornith-1.5-9B, fused using Adaptive Geodesic Spectral Interpolation (AGSI). This model excels at dense mathematical deduction, programmatic problem-solving, long-horizon chain-of-thought reasoning, and agentic capabilities like terminal/CLI mastery and autonomous bash execution, with a context length of 131,072 tokens.

Loading preview...

MiMo-Ornith-9B-AGSI: Advanced Agentic & Reasoning Model

MiMo-Ornith-9B-AGSI, developed by OliviaRossi, is a 9 billion parameter model built upon the Qwen 9B architecture. It represents a novel non-linear synthesis of two distinct, highly specialized models:

  • XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B: Optimized for dense mathematical deduction, SWE-bench verified programmatic problem-solving, and long-horizon chain-of-thought (CoT) reasoning.
  • ornith-ai/Ornith-1.5-9B: Specializes in agentic capabilities, including terminal/CLI mastery, autonomous bash execution, self-debugging loops, and algorithmic generation.

Key Differentiator: Adaptive Geodesic Spectral Interpolation (AGSI)

Unlike traditional linear or spherical interpolation methods, MiMo-Ornith-9B-AGSI was fused using Adaptive Geodesic Spectral Interpolation (AGSI). This advanced technique addresses common issues like Frobenius Attenuation, Isotropic Collapsing, and Anti-Phase Annihilation by:

  • Direction-Magnitude (DoRA) Decoupling: Isolates radial feature scale from directional orientation.
  • Row-Wise Hyperspherical Geodesics: Calculates angular geodesic distance for each neuron row.
  • Anti-Phase Interference Shielding: Prevents destructive cancellation when gradients pull in opposing directions.
  • Quadratic Spectral Energy Invariant: Maintains stable forward-pass activations by preserving parameter norm.
  • Quintic Smoothstep Depth & Functional Block Routing: Dynamically adjusts mixing coefficients across transformer layers, biasing specific layers towards MiMo for syntactic anchoring and factual retention, and towards Ornith for action policy and tool-calling.

Capabilities & Use Cases

This model is particularly well-suited for complex technical tasks requiring both deep reasoning and autonomous execution. Its capabilities include:

  • Advanced Problem Solving: Excels in mathematical deduction and programmatic problem-solving.
  • Agentic Functionality: Proficient in terminal/CLI interactions, bash execution, and self-debugging.
  • Code Generation: Capable of generating complex algorithms and production-ready code, including eBPF programs.
  • Long-Context Reasoning: Supports a substantial context length of 131,072 tokens, enabling intricate chain-of-thought processes.

It is compatible with standard Qwen 9B architecture systems like vLLM and transformers, and uses a unified Qwen chat template for prompt formatting, separating internal reasoning steps with <think> tags.