Qwen2.5-7B-Instruct-mlx-8bit

8-bit MLX conversion of Qwen/Qwen2.5-7B-Instruct optimized for Apple Silicon native GPU inference.

Converted by: SirSahOl
Source Model: Qwen/Qwen2.5-7B-Instruct
Framework: MLX by Apple
Quantization: 8-bit (Average 8.25 bits per weight)
Format: safetensors
License: apache-2.0

---

Model Details

• Architecture: Qwen2ForCausalLM
• Parameters: 7.61B (7.07B non-embedding)
• Context Length: 32,768 tokens (native, extensible up to 131,072 with YaRN)
• Format: MLX (Apple Silicon native GPU format)
• Quantization: 8-bit (Average 8.25 bits per weight)
• Active VRAM Footprint: ~7.8 GB (Minimum recommended: 16 GB Unified Memory)

---

Quick Start

Installation

pip install mlx-lm

CLI Usage

# Chat interactively
mlx_lm.chat --model SirSahOl/Qwen2.5-7B-Instruct-chat-mlx-8bit

# Generate text
mlx_lm.generate --model SirSahOl/Qwen2.5-7B-Instruct-chat-mlx-8bit --prompt "Write a short poem about artificial intelligence."

Python API (with Chat Template)

from mlx_lm import load, generate

model, tokenizer = load("SirSahOl/Qwen2.5-7B-Instruct-chat-mlx-8bit")

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:

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:

---

Who Should Use This?

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

---

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 Qwen2.5.json in your LM Studio config directory:

• macOS / Linux: ~/.lmstudio/config-presets/Qwen2.5.json
• Windows: %USERPROFILE%\.lmstudio\config-presets\Qwen2.5.json

{
  "name": "Qwen2.5",
  "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/Qwen2.5-7B-Instruct-chat-mlx-8bit, and select "Qwen2.5" from the Prompt Template dropdown.

Option B: Manual LM Studio Configuration

1. Load SirSahOl/Qwen2.5-7B-Instruct-chat-mlx-8bit 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:

FROM SirSahOl/Qwen2.5-7B-Instruct-chat-mlx-8bit

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:

ollama create qwen2.5-7b -f Modelfile
ollama run qwen2.5-7b

---

Conversion Details

Reproduction

To reproduce this conversion:

pip install mlx-lm==0.31.3
python3 -m mlx_lm.convert --hf-path Qwen/Qwen2.5-7B-Instruct --mlx-path output/Qwen2.5-7B-Instruct-mlx-8bit -q --q-bits 8

---

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 for full license details.

---

Changelog

---

Converted with MLX Foundry — a professional pipeline for converting models to Apple MLX format.