---
library_name: mlx
pipeline_tag: text-generation
license: apache-2.0
base_model: TokenRhythm/NeoHorse-1-9B
base_model_relation: quantized
tags:
  - mlx
  - safetensors
  - apple-silicon
  - 4-bit
  - conversational
  - transformers
  - qwen3_5_text
  - text-generation
  - agentic
  - tool-use
  - coding
  - reasoning
  - instruction-following
  - arxiv:2609.08183
  - license:apache-2.0
  - endpoints_compatible
---

# NeoHorse-1-9B-mlx-4bit

> 4-bit MLX conversion of [TokenRhythm/NeoHorse-1-9B](https://huggingface.co/TokenRhythm/NeoHorse-1-9B) optimized for Apple Silicon native GPU inference.

**Converted by**: [SirSahOl](https://huggingface.co/SirSahOl)  
**Source Model**: [TokenRhythm/NeoHorse-1-9B](https://huggingface.co/TokenRhythm/NeoHorse-1-9B)  
**Framework**: [MLX](https://github.com/ml-explore/mlx) by Apple  
**Quantization**: 4-bit (Average 4.50 bits per weight)  
**Format**: safetensors  
**License**: apache-2.0  

---

## Model Details

*   **Architecture**: Qwen3_5ForCausalLM
*   **Parameters**: 9.0B
*   **Context Length**: 262,144 tokens
*   **Format**: MLX (Apple Silicon native GPU format)
*   **Quantization**: 4-bit (Average 4.50 bits per weight)
*   **Active VRAM Footprint**: ~5.4 GB (Minimum recommended: 16 GB Unified Memory)

---

## Quick Start

### Installation

```bash
pip install mlx-lm
```

### CLI Usage

```bash
# Chat interactively
mlx_lm.chat --model SirSahOl/NeoHorse-1-9B-chat-mlx-4bit

# Generate text
mlx_lm.generate --model SirSahOl/NeoHorse-1-9B-chat-mlx-4bit --prompt "Write a short poem about artificial intelligence."
```

### Python API (with Chat Template)

```python
from mlx_lm import load, generate

model, tokenizer = load("SirSahOl/NeoHorse-1-9B-chat-mlx-4bit")

messages = [
    {"role": "user", "content": "Explain quantum superposition in simple terms."}
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)
```

---

## Performance Benchmarks

### Apple Silicon Hardware Sizing Matrix

Estimated decoding throughput, time-to-first-token (TTFT), and active unified memory footprint across Apple Silicon tiers:

| Apple Silicon Tier | Unified Memory | Active VRAM | Estimated Speed | Est. TTFT | Recommended Use Case |
| :--- | :--- | :--- | :--- | :--- | :--- |
| **M1 / M2 / M3 / M4 (Base)** | 16 GB Unified Memory | ~5.4 GB | **~28 tokens/sec** | ~141 ms | Everyday interactive assistant & fast local completions |
| **M1 / M2 / M3 / M4 Pro** | 18 GB – 36 GB | ~5.4 GB | **~42 tokens/sec** | ~96 ms | Balanced daily driver for coding, tool invocation, and multi-turn chat |
| **M1 / M2 / M3 / M4 Max** | 36 GB – 128 GB | ~5.4 GB | **~60 tokens/sec** | ~59 ms | High-throughput generation, low latency, agent orchestration |
| **M1 / M2 / M3 Ultra** | 64 GB – 192 GB | ~5.4 GB | **~84 tokens/sec** | ~39 ms | Peak concurrency, batch document extraction, production serving |

> *Estimates based on Apple Silicon unified memory bandwidth and active parameter footprint. Real-world speeds may vary with context length.*

---

## Multi-Quantization Comparison

Evaluate your hardware budget and choose the optimal precision:

| Variant | Disk Size | VRAM Footprint | Target Apple Silicon Hardware | Key Advantage |
| :--- | :--- | :--- | :--- | :--- |
| **4-bit MLX** (This Repository) | ~5.3 GB | ~5.3 GB | M1 / M2 / M3 / M4 (8GB+) | Maximum generation speed and lowest RAM overhead. |
| **[8-bit MLX](https://huggingface.co/SirSahOl/NeoHorse-1-9B-chat-mlx-8bit)** | ~9.9 GB | ~9.9 GB | M1 / M2 / M3 / M4 Pro/Max (16GB+) | Balanced accuracy and generation speed; near-lossless reasoning. |
| **[16-bit MLX](https://huggingface.co/SirSahOl/NeoHorse-1-9B-chat-mlx-16bit)** | ~18.8 GB | ~18.8 GB | M2 / M3 / M4 Max/Ultra (32GB+) | Full unquantized precision; reference evaluation quality. |

---

## Who Should Use This?

| Your Hardware | Recommended Quantization |
|---------------|--------------------------|
| M1/M2/M3/M4 (8GB – 16GB) | **4-bit** — Best balance of speed, low memory, and multitasking capability |
| M1/M2/M3/M4 Pro/Max (18GB – 36GB) | **8-bit** — Higher quality reasoning with comfortable memory headroom |
| M1/M2/M3/M4 Max/Ultra (36GB – 192GB) | **16-bit** — Unquantized full precision, zero quality degradation |

**General guidance:**
- Use **4-bit** if you want to run this model alongside IDEs, browsers, and background development tools.
- Use **8-bit** if you have 16GB+ unified memory and require superior reasoning and code accuracy.
- Use **16-bit** for research, benchmarking, evaluation, or high-end workstation deployments.

---

## Other Quantization Variants

| Variant | Link |
|---------|------|
| 4-bit | [SirSahOl/NeoHorse-1-9B-chat-mlx-4bit](https://huggingface.co/SirSahOl/NeoHorse-1-9B-chat-mlx-4bit) |
| 8-bit | [SirSahOl/NeoHorse-1-9B-chat-mlx-8bit](https://huggingface.co/SirSahOl/NeoHorse-1-9B-chat-mlx-8bit) |
| 16-bit | [SirSahOl/NeoHorse-1-9B-chat-mlx-16bit](https://huggingface.co/SirSahOl/NeoHorse-1-9B-chat-mlx-16bit) |

---

## LM Studio & Local Inference Setup Guide

To prevent runaway loops and ensure correct conversational turn-taking, configure custom stop strings in your local inference runtime.

### Custom Stop Strings Configuration

Ensure the following sequence tokens are configured as strict stop boundaries:

1. `<|im_start|>`
2. `<|im_end|>`
3. `<|endoftext|>`

### Option A: Automatic Preset (Recommended for LM Studio)

Create a custom prompt preset JSON file named `ChatML.json` in your LM Studio config directory:
- **macOS / Linux**: `~/.lmstudio/config-presets/ChatML.json`
- **Windows**: `%USERPROFILE%\.lmstudio\config-presets\ChatML.json`

```json
{
  "name": "ChatML",
  "inference_params": {
    "pre_prompt": "You are a helpful, respectful, and honest assistant.",
    "input_prefix": "<|im_start|>user\n",
    "input_suffix": "<|im_end|>\n",
    "pre_prompt_prefix": "<|im_start|>system\n",
    "pre_prompt_suffix": "<|im_end|>\n",
    "antiprompt": [
      "<|im_start|>",
      "<|im_end|>",
      "<|endoftext|>"
    ],
    "stopStrings": [
      "<|im_start|>",
      "<|im_end|>",
      "<|endoftext|>"
    ],
    "temperature": 0.7,
    "max_tokens": 2048
  }
}
```

Restart LM Studio, load **SirSahOl/NeoHorse-1-9B-chat-mlx-4bit**, and select **"ChatML"** from the Prompt Template dropdown.

### Option B: Manual LM Studio Configuration

1. Load **SirSahOl/NeoHorse-1-9B-chat-mlx-4bit** in LM Studio.
2. In the right-hand **Advanced Configuration / Inference Parameters** panel:
   - **Stop Strings (antiprompt / stopStrings)**: Add `<|im_start|>`, `<|im_end|>`, `<|endoftext|>`
   - **System Prefix**: `<|im_start|>system\n`
   - **System Suffix**: `<|im_end|>\n`
   - **User Prefix**: `<|im_start|>user\n`
   - **Assistant Suffix**: `<|im_end|>\n<|im_start|>assistant\n`

### Option C: Ollama Modelfile Setup

Create a `Modelfile` to run this model in Ollama:

```dockerfile
FROM SirSahOl/NeoHorse-1-9B-chat-mlx-4bit

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
```

Create and run with Ollama:
```bash
ollama create neohorse-1-9b-chat-mlx-4bit -f Modelfile
ollama run neohorse-1-9b-chat-mlx-4bit
```

---

## Conversion Details

| Property | Value |
|----------|-------|
| **Source Model** | [TokenRhythm/NeoHorse-1-9B](https://huggingface.co/TokenRhythm/NeoHorse-1-9B) |
| **Quantization** | 4-bit |
| **mlx-lm Version** | 0.31.3 |
| **Conversion Time** | 101.37s |
| **Output Size** | 4.7 GB |
| **Date** | 2026-09-14T23:31:53.987005+00:00 |

### Reproduction

To reproduce this conversion:

```bash
pip install mlx-lm==0.31.3
python3 -m mlx_lm.convert --hf-path /root/.cache/huggingface/hub/models--TokenRhythm--NeoHorse-1-9B/snapshots/ba5b6e40d88a6ddf4591e176738254a3bc715765 --mlx-path output/NeoHorse-1-9B-mlx-4bit -q --q-bits 4
```

---

## Limitations & Known Issues

- 4-bit group-wise quantization introduces minor precision loss compared to unquantized weights; for deep mathematical derivations or precision-critical reasoning, test the 8-bit or 16-bit variants.
- High context sequences (>32K tokens) require sufficient unified memory headroom; ensure unified memory is not overcommitted.
- This is a weight-only MLX conversion designed specifically for Apple Silicon GPUs (M1/M2/M3/M4 series).

---

## License

This model conversion inherits the license of the source model: **apache-2.0**.

See the [original model card](https://huggingface.co/TokenRhythm/NeoHorse-1-9B) for full license details.

---

## Changelog

| Version | Date | Changes |
|---------|------|---------|
| v1.0 | 2026-09-14 | Initial conversion |

---

*Converted with [MLX Foundry](https://github.com/SirSahOl/mlx-foundry) — a professional pipeline for converting models to Apple MLX format.*