Ornith-1.0-9B-oQe3.5

<p align="center"> <img src="ornith_logo.png" alt="Ornith-1.0-9B" width="640"/> </p>

Ornith-1.0-9B-oQe3.5

Apple Silicon Optimized oQe3.5 MLX Quantized Release

This repository contains an oQe3.5 mixed-precision MLX quantized version of Ornith-1.0-9B, optimized for fast and memory-efficient inference on Apple Silicon.

The original Ornith-1.0-9B model was developed by DeepReinforce. This repository provides an optimized MLX conversion only and does not contain any additional fine-tuning or retraining. The goal is to preserve the original model's exceptional coding and reasoning performance while significantly reducing memory requirements through oQe3.5 sensitivity-aware mixed-precision quantization. :contentReference[oaicite:0]{index=0}

---

About Ornith-1.0

Ornith-1.0 is a family of open-source reasoning models specialized for agentic software engineering and coding workflows.

The 9B model is the most compact member of the Ornith family, delivering strong performance while remaining practical for local deployment.

Key capabilities include:

• 🧠 Advanced reasoning
• 💻 Agentic coding
• 🛠 Native tool calling
• 🔧 Function calling
• 📚 Long-context reasoning
• 🤖 Multi-step planning
• 🔍 Codebase understanding
• ⚡ Software engineering automation

Ornith is trained using a reinforcement learning framework that jointly optimizes both solution generation and the reasoning scaffolds used to reach those solutions, enabling stronger search trajectories and higher-quality coding performance. :contentReference[oaicite:1]{index=1}

---

Quantization

This release uses oQe3.5 mixed-precision quantization.

Specifications

• Format: MLX
• Quantization: oQe3.5
• Method: Sensitivity-Aware Mixed Precision
• Platform: Apple Silicon
• Inference Engine: MLX / oMLX

Unlike traditional fixed-bit quantization, oQe3.5 dynamically assigns precision based on layer sensitivity, preserving higher precision where it matters most while aggressively compressing less sensitive layers.

Benefits include:

• Higher reasoning quality
• Better coding accuracy
• Lower memory usage
• Faster inference
• Excellent Apple Silicon efficiency

---

Recommended Settings

For the best overall performance:

temp: 0.6
top_p: 0.95
top_k: 20
min_p: 0
rep_penalty: 1.05
presence_penalty: 1.2
enable_thinking: true

For benchmark reproduction:

temp: 1.0
top_p: 0.95
top_k: 20

These settings align with the recommendations from the original Ornith model card. :contentReference[oaicite:2]{index=2}

---

Example Usage

from mlx_lm import load, generate

model, tokenizer = load("yugeshkarunamurthy/Ornith-1.0-9B-oQe3.5")

messages = [
    {
        "role": "user",
        "content": "Write a Python implementation of an LRU cache."
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    temp=0.6,
    top_p=0.95,
    top_k=20,
    max_tokens=16384,
)

print(response)

---

Optimized For

This release is optimized for:

• Apple M1
• Apple M2
• Apple M3
• Apple M4

Compatible with:

• MLX
• MLX-LM
• oMLX
• Open WebUI
• LM Studio (MLX)
• Local AI Agent Frameworks

---

Model Highlights

• State-of-the-art open coding model in the 9B class
• Advanced reasoning
• Agentic software engineering
• Native function calling
• Native tool use
• Long-context reasoning
• Multi-step planning
• Repository understanding
• Research assistance

---

Intended Use

This model is particularly well suited for:

• Software Engineering
• AI Coding Assistants
• Autonomous Coding Agents
• Repository Analysis
• Bug Fixing
• Code Generation
• Research Automation
• Multi-step Planning
• Local AI Development

---

Hardware Recommendations

Recommended systems:

• Apple M1 Pro / Max / Ultra
• Apple M2 Pro / Max / Ultra
• Apple M3 Series
• Apple M4 Series

Higher-memory Apple Silicon systems provide the best experience for long-context coding sessions.

---

About oQe3.5 Quantization

oQe3.5 is a sensitivity-aware mixed-precision quantization method developed to maximize model quality while substantially reducing memory requirements.

Instead of assigning the same precision to every weight matrix, oQe3.5 automatically allocates precision according to each layer's importance.

This approach enables:

• Better reasoning preservation
• Higher coding quality
• Faster inference
• Lower RAM usage
• Excellent Apple Silicon performance

---

Original Model

The original Ornith-1.0-9B introduces a self-improving reinforcement learning framework for agentic coding and is released under the MIT License.

Notable features include:

• Reinforcement Learning for agentic reasoning
• Tool-aware reasoning
• Native XML tool calling
• Coding-first optimization
• 262K context window
• Open-source MIT license

For detailed benchmarks, evaluation methodology, and technical documentation, please visit the original model page. :contentReference[oaicite:3]{index=3}

---

Credits

Original Model

All credit for the original model, datasets, training methodology, evaluation, benchmarks, and research belongs entirely to:

DeepReinforce Team

Original Repository:

https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B

Project Page:

https://deep-reinforce.com/ornith_1_0.html

---

oQe3.5 MLX Quantized Release

This repository provides an Apple Silicon optimized oQe3.5 MLX quantized version of the original model.

No additional fine-tuning has been performed.

---

Acknowledgements

• DeepReinforce
• Qwen Team
• Apple MLX
• Hugging Face
• Transformers
• vLLM
• SGLang
• oMLX
• OptiQ Quantization

---

Citation

If you use this model in research, please cite the original Ornith paper:

@misc{ornith_9b,
    title={{Ornith-1.0-9B}: Agentic Coding, Open to All},
    author={DeepReinforce Team},
    year={2026},
    url={https://deep-reinforce.com/ornith_1_0.html}
}

---

License

This release inherits the MIT License from the original model.

Please refer to the original repository for complete licensing information.

---

Disclaimer

This repository contains an optimized oQe3.5 MLX quantized conversion intended for efficient local inference on Apple Silicon.

All original model architecture, datasets, training methodology, benchmarks, evaluations, and research remain entirely the work of the original DeepReinforce team.