---
license: mit
license_link: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B
base_model:
- deepreinforce-ai/Ornith-1.0-9B
base_model_relation: quantized
quantized_by: AtomicChat
pipeline_tag: text-generation
library_name: mlx
tags:
- atomic-chat
- ornith
- deepreinforce
- coding
- agent
- mlx
- apple-silicon
- quantized
- qwen3_5
---
**Ornith 1.0 9B**, quantized to **MLX 4-bit** by [Atomic Chat](https://atomic.chat) for Apple Silicon. Built straight from DeepReinforce's original weights. Runs fully offline on your Mac.
## Highlights
- **A self-improving open-source family for agentic coding** from DeepReinforce, built for tool-calling and terminal-based coding agents.
- **Post-trained on top of Gemma 4 and Qwen 3.5**, the smallest, fastest member of the Ornith 1.0 lineup.
- **Strong agentic coding scores for its size**: 69.4 on SWE-bench Verified and 43.1 on Terminal-Bench 2.1 (Terminus-2).
- **Dense architecture, 32 layers**, `qwen3_5` model type with a `hidden_size` of 4096.
- **262,144-token native context** for long files and multi-step agent traces.
- **Pure open**: MIT licensed, globally accessible with no regional limits.
- **Full quant ladder** with an importance matrix on every quant over [`calibration_datav3`](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8).
> [!NOTE]
> This is the **MLX 4-bit** build for Apple Silicon (M-series). For llama.cpp/Ollama/CPU use the [GGUF repo](https://huggingface.co/AtomicChat/ornith-9b-GGUF).
## Model Overview
| Property | Value |
|---|---|
| Base model | `deepreinforce-ai/Ornith-1.0-9B` |
| Total parameters | ~9B (model name; card states no exact figure in prose) |
| Layers | 32 |
| Context length | 262,144 |
| Architecture | `qwen3_5` dense causal LM, post-trained on Gemma 4 and Qwen 3.5 |
| This repo | MLX **4-bit** quant for Apple Silicon (~5.0 GB), built from the original weights. |
Scores are DeepReinforce's published results for the full-precision base `deepreinforce-ai/Ornith-1.0-9B`. MLX quants run the same model locally; lower bit-widths trade a little accuracy for size/speed.
## MLX quants in this series
[4-bit](https://huggingface.co/AtomicChat/ornith-9b-MLX-4bit) ← this · [5-bit](https://huggingface.co/AtomicChat/ornith-9b-MLX-5bit) · [6-bit](https://huggingface.co/AtomicChat/ornith-9b-MLX-6bit) · [8-bit](https://huggingface.co/AtomicChat/ornith-9b-MLX-8bit)
## Run on Apple Silicon
```bash
pip install mlx-lm
mlx_lm.generate --model AtomicChat/ornith-9b-MLX-4bit --prompt "Write a quicksort in Python" --max-tokens 512
```
```python
from mlx_lm import load, generate
model, tokenizer = load("AtomicChat/ornith-9b-MLX-4bit")
msg = [{"role": "user", "content": "Write a quicksort in Python"}]
prompt = tokenizer.apply_chat_template(msg, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512, verbose=True))
```
Or open it in **[Atomic Chat](https://atomic.chat)**: search `AtomicChat/ornith-9b-MLX-4bit` and hit **Use this model**.
## Recommended sampling
| Parameter | Value |
|---|---|
| temperature | 0.6 |
| top_p | 0.95 |
| top_k | 20 |
DeepReinforce's recommended sampling parameters. The card notes that `temperature=1.0` reproduces the reported benchmark setup.
## How this was made
1. Download `deepreinforce-ai/Ornith-1.0-9B` (original weights).
2. Convert + quantize to MLX with `mlx_lm.convert -q --q-bits 4 --q-group-size 64`.
## License
Released by DeepReinforce under the MIT license, globally accessible with no regional limits. Quantized to MLX by Atomic Chat.
