---

license: mit

language:

• multilingual

base_model:

• ornith-ai/Ornith-1.5-35B-A3B

tags:

• large-language-model
• reasoning
• mixture-of-experts
• agentic-coding
• code-generation
• tool-use
• long-context

---

Ornith-1.5-35B-A3B

Ornith-1.5-35B-A3B is a mixture-of-experts reasoning model from the Ornith family, designed for demanding software-engineering and agentic workloads. With approximately 35 billion total parameters and around 3 billion active parameters per token, the model combines a large parameter pool with a comparatively small active computation footprint.

This repository provides GGUF quantized variants of Ornith-1.5-35B-A3B for efficient local inference. The available quantization formats substantially reduce storage and memory requirements while retaining the model's reasoning, coding, and agent-oriented capabilities.

Ornith-1.5-35B-A3B is intended for multi-step development workflows where the model may need to understand existing code, reason through problems, generate or modify implementation details, and interact with tools through a compatible inference or agent framework.

---

Model Overview

• Model Name: Ornith-1.5-35B-A3B
• Base Model: ornith-ai/Ornith-1.5-35B-A3B
• Architecture: Mixture-of-Experts language model
• Parameter Count: Approximately 35 Billion total parameters
• Active Parameters: Approximately 3 Billion parameters per token
• Model Type: Reasoning
• Modalities: Text
• Context Length: 262,144 tokens
• Developer: Ornith AI
• License: MIT

---

Quantization Formats

This repository provides multiple GGUF quantized versions of Ornith-1.5-35B-A3B. Each format offers a different trade-off between model storage, memory consumption, inference requirements, and preservation of the original model's behavior.

IQ3_M

• Size reduction of approx 79.53% (14.72 GB) compared to 16-bit (71.90 GB)
• Aggressive 3-bit quantization intended for substantially lower memory requirements
• Suitable for local agentic coding and reasoning workloads where deployment resources are constrained
• Makes the large 35B-parameter MoE model considerably more practical for local inference
• Fine-grained reasoning, complex code generation, and long multi-step workflows may experience greater degradation than higher-precision variants

IQ4_NL

• Size reduction of approx 73.55% (19.02 GB) compared to 16-bit (71.90 GB)
• 4-bit non-linear quantization providing a stronger balance between model footprint and output fidelity
• Suitable for coding agents, software-engineering tasks, repository analysis, and reasoning workloads
• Retains more representational capacity than the more aggressive IQ3_M format while remaining substantially smaller than the F16 model
• Appropriate for deployments where memory efficiency and reasoning quality both need to be considered

IQ4_XS

• Size reduction of approx 74.88% (18.06 GB) compared to 16-bit (71.90 GB)
• Compact 4-bit quantization designed for efficient local deployment of the MoE reasoning model
• Provides a practical compromise between memory consumption and retention of coding and reasoning behavior
• Suitable for software-development assistants, agentic workflows, and multi-step problem solving
• Offers a substantially smaller deployment footprint than the original 16-bit representation

---

Training Background (Original Model)

Ornith-1.5-35B-A3B belongs to the Ornith model family focused on reasoning and agentic software engineering. Its mixture-of-experts architecture separates total model capacity from the amount of computation activated for each token, allowing the model to maintain a large parameter pool while using only a subset of parameters during individual forward passes.

Mixture-of-Experts Architecture

• Uses a sparse MoE architecture with approximately 35B total parameters
• Approximately 3B parameters are active for each token
• The sparse activation pattern reduces per-token computational requirements relative to a dense model containing the same total parameter count
• Provides a large parameter capacity for reasoning and coding while maintaining a comparatively smaller active computation footprint

---

Key Capabilities

• Agentic Coding

Designed for software-engineering workflows involving iterative reasoning, implementation, debugging, and interaction with development tools.

• Code Generation

Generates implementation code for software-development tasks while reasoning about the surrounding problem context.

• Code Modification

Can be used in workflows that require modifying existing implementations rather than generating standalone snippets.

• Repository-Level Reasoning

Supports analysis of larger software projects and tasks involving relationships between multiple source files.

• Complex Reasoning

Designed for multi-step problems where solving the task requires planning, intermediate reasoning, and iterative refinement.

• Tool Use

Can be integrated with compatible agentic runtimes that provide external tools, APIs, terminals, or development environments.

• Mixture-of-Experts Efficiency

The 35B total parameter capacity is combined with approximately 3B active parameters per token, providing sparse computation during inference.

• Long-Context Workflows

Intended for extended interactions and software-engineering tasks where substantial contextual information may need to remain available.

---

Recommended Usecases

• Agentic Coding Systems

Build coding agents capable of reasoning through software-development tasks and interacting with external development tools.

• Repository-Level Software Engineering

Analyze and work with larger codebases containing multiple files, dependencies, and implementation components.

• Automated Debugging

Assist with investigating software problems, identifying likely causes, and developing implementation-level fixes.

• Code Generation and Refactoring

Generate new implementations or assist with restructuring and improving existing software components.

• Terminal-Based Coding Agents

Integrate the model with terminal environments and agent frameworks for execution-oriented software development.

• Developer Productivity

Automate repetitive development tasks such as code generation, modification, analysis, and debugging assistance.

• Tool-Calling Applications

Connect the model to external functions, APIs, development tools, or agent orchestration frameworks.

• Research and Experimentation

Evaluate sparse MoE reasoning models, quantization techniques, local inference, and agentic coding workflows.

---

Usage Example

Using llama.cpp

./llama-cli \
  -m SandLogicTechnologies/Ornith-1.5-35B-IQ3_M.gguf \
  -mmproj SandLogicTechnologies/mmproj-Ornith-1.5-35B-A3B-bf16.gguf

Acknowledgments

These GGUF quantized models are based on the original work of the Ornith AI team.

Special thanks to:

• The Ornith AI team for developing and releasing the Ornith-1.5-35B-A3B model.

• The developers and contributors of llama.cpp and the broader GGUF ecosystem for enabling efficient local inference and quantized model deployment.

---

Contact

For questions, feedback, or support, please refer to the official Ornith AI project and model repository.