Ornith-1.0-9B-GGUF

Ornith-1.0-9B

Ornith-1.0-9B is a dense 9B-parameter reasoning model developed by the Ornith/DeepReinforce team as part of the Ornith-1.0 family of open-source models specialized for agentic coding. It is designed to operate in multi-step software engineering workflows, combining reasoning, tool calling, code generation, and iterative problem solving.

This repository contains GGUF quantized variants of Ornith-1.0-9B optimized for efficient local inference. The quantized formats substantially reduce the memory footprint of the model while retaining its core reasoning and agentic coding capabilities.

Ornith-1.0-9B is a reasoning model that produces an explicit reasoning phase before its final response and supports structured tool calls when deployed through compatible serving frameworks. The official model card reports strong results on agentic coding benchmarks including SWE-bench Verified, SWE-bench Pro, Terminal-Bench 2.1, NL2Repo, and ClawEval.

---

Model Overview

• Model Name: Ornith-1.0-9B
• Base Model: ornith-ai/Ornith-1.0-9B
• Architecture: Dense language model
• Parameter Count: Approximately 9 Billion Parameters
• Modalities: Text, Image
• Model Type: Reasoning / Agentic Coding Model
• Context Length: Up to 262,144 tokens in the official serving configurations
• Developer: Ornith / DeepReinforce Team
• License: MIT

---

Quantization Formats

This repository provides various GGUF quantized versions of the Ornith-1.0-9B model optimized for efficient local inference.

IQ3_M

• Size reduction of approx 75.37% (4.11 GB)compared to 16-bit (16.69 GB)
• Aggressive 3-bit quantization designed to minimize memory requirements for agentic reasoning workloads
• Suitable for local coding agents, developer tools, and constrained inference environments
• Enables deployment of a 9B reasoning model on substantially smaller hardware configurations
• Complex software-engineering reasoning, long-context interactions, and multi-step tool workflows may experience greater quality degradation than higher-precision variants

IQ4_NL

• Size reduction of approx 69.62% (5.07 GB)compared to 16-bit (16.69 GB)
• Advanced 4-bit non-linear quantization designed to retain reasoning and code-generation fidelity
• Better suited for agentic coding, debugging, repository-level analysis, and tool-assisted workflows
• Provides a stronger balance between model capability and memory consumption than the more aggressive IQ3_M format
• May introduce additional computational overhead during inference compared with lower-bit quantization

IQ4_XS

• Size reduction of approx 70.82% (4.87 GB)compared to 16-bit (16.69 GB)
• Balanced 4-bit quantization focused on maintaining reasoning stability while reducing deployment requirements
• Suitable for local coding agents, software-engineering assistants, and extended problem-solving sessions
• Provides a practical compromise between memory footprint, generation quality, and inference efficiency
• Maintains dependable behavior across most practical agentic coding workloads

Q6_K

• Size reduction of approx 58.96% (6.85 GB)compared to 16-bit (16.69 GB)
• Higher-precision 6-bit K-Quant format intended to preserve more of the original model's reasoning and generation behavior
• Better suited for demanding coding-agent workflows, repository analysis, complex debugging, and multi-step problem solving
• Provides greater headroom for retaining model fidelity than the lower-bit variants
• Recommended when reasoning quality and coding reliability are more important than minimizing memory consumption

---

Training Background (Original Model)

Ornith-1.0 is described by its developers as a self-improving family of open-source models specialized for agentic coding. The training approach uses reinforcement learning to optimize not only solution rollouts but also the scaffolds that drive those rollouts. This allows the model to learn search trajectories for solving complex software-engineering tasks.

Agentic Training

• Reinforcement-learning-based training focused on agentic software-engineering workflows
• Joint optimization of solution rollouts and the scaffolds used to produce those solutions
• Training designed to improve search trajectories and multi-step problem solving
• Post-training performed on top of model families including Gemma 4 and Qwen 3.5 within the broader Ornith-1.0 family

Reasoning and Tool Use

• Generates a reasoning block before the final answer
• Supports structured tool calls through compatible serving runtimes
• Designed to interact with coding environments and external tools
• Optimized for terminal-oriented and repository-level software engineering workflows

---

Key Capabilities

• Agentic Coding Designed to solve software-engineering tasks through iterative reasoning, tool interaction, and execution-oriented workflows.

• Code Generation & Modification Generates and modifies code as part of larger software-development tasks rather than limiting itself to isolated code snippets.

• Repository-Level Reasoning Supports analysis of larger software projects and multi-file development problems.

• Tool Calling Produces structured function calls that can be surfaced through OpenAI-compatible serving APIs.

• Reasoning Uses an explicit reasoning stage to work through complex multi-step problems before producing the final response.

• Long-Context Processing The official serving configurations support contexts up to 262,144 tokens, enabling analysis of large codebases and extended agent interactions.

• Multimodal Input The Hugging Face model is exposed with image-text-to-text support, allowing visual information to be incorporated alongside textual prompts.

---

Recommended Usecases

• Agentic Coding Systems Build coding agents capable of reasoning through software tasks and interacting with development tools.

• Repository-Level Software Engineering Analyze and modify larger codebases involving multiple files and dependencies.

• Automated Debugging Investigate software problems through iterative reasoning and tool-assisted workflows.

• Terminal-Based Coding Agents Integrate the model into command-line development environments and coding-agent frameworks.

• Developer Productivity Automate repetitive software-engineering tasks, code changes, and repository operations.

• Tool-Calling Applications Connect the model to external functions, APIs, shell tools, or agent frameworks.

• Research & Experimentation Evaluate reinforcement-learning-based agentic coding models and local reasoning systems.

---

Usage Example

Using llama.cpp

./llama-server \
  -hf SandLogicTechnologies/Ornith-1.0-9B-GGUF \
  --port 8000 \
  -c 262144

For the GGUF files in this repository, select the desired quantization according to available memory and required reasoning fidelity.

Acknowledgments

These quantized models are based on the original work of the Ornith / DeepReinforce team.

Special thanks to:

• The Ornith / DeepReinforce team for developing and releasing Ornith-1.0-9B.

• The developers and open-source community behind llama.cpp for enabling efficient GGUF-based inference.

---

Contact

For questions, feedback, or support, please reach out at [email protected] or visit https://www.sandlogic.com/