AITRADER/Amsi-fin-o1-MLX-8bit

Amsi-fin-o1 MLX (8-bit Quantized)

Quantized Financial Vision-Language Model for Apple Silicon

Model Description

This is the 8-bit quantized MLX conversion of AITRADER/Amsi-fin-o1, a specialized financial vision-language model. The 8-bit quantization reduces memory usage by ~50% while maintaining excellent performance for financial analysis tasks.

Key Features

• Reduced Memory Footprint: ~4GB vs ~8GB for bf16 version
• Faster Inference: Optimized for speed on Apple Silicon
• Financial Document Analysis: Extract and analyze financial statements and reports
• Chart Understanding: Interpret financial charts and visualizations
• Chain-of-Thought Reasoning: Advanced thinking for complex financial calculations
• Apple Silicon Native: Optimized for M1/M2/M3/M4 chips

Quantization Details

Parameter	Value
Quantization Type	8-bit Integer
Group Size	64
Memory Reduction	~50%
Quality Retention	~98%+

Model Architecture

Component	Specification
Base Architecture	Qwen3-VL (4B parameters)
Text Model	36 layers, 2560 hidden size
Vision Encoder	24 layers, 1024 hidden size
Attention Heads	32 (8 KV heads)
Context Length	Up to 131,072 tokens
Precision	8-bit Quantized
Model Size	~4GB

Installation

# Install mlx-vlm
pip install -U mlx-vlm

Quick Start

Command Line

# Basic image analysis
python -m mlx_vlm.generate \
    --model AITRADER/Amsi-fin-o1-MLX-8bit \
    --max-tokens 512 \
    --temperature 0.7 \
    --prompt "Analyze this financial chart and explain the trends." \
    --image path/to/financial_chart.png

Python API

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load model
model_path = "AITRADER/Amsi-fin-o1-MLX-8bit"
model, processor = load(model_path)
config = load_config(model_path)

# Prepare prompt
prompt = apply_chat_template(
    processor,
    config,
    "Analyze the financial performance shown in this quarterly report.",
    num_images=1
)

# Generate response
output = generate(
    model,
    processor,
    prompt,
    image="path/to/report.png",
    max_tokens=512,
    temperature=0.7
)
print(output)

Streaming Generation

from mlx_vlm import load, stream_generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "AITRADER/Amsi-fin-o1-MLX-8bit"
model, processor = load(model_path)
config = load_config(model_path)

prompt = apply_chat_template(
    processor,
    config,
    "What are the key insights from this financial statement?",
    num_images=1
)

# Stream the response
for token in stream_generate(
    model,
    processor,
    prompt,
    image="financial_statement.png",
    max_tokens=512
):
    print(token, end="", flush=True)

Use Cases

1. Financial Chart Analysis

prompt = """Analyze this stock chart and provide:
1. The overall trend (bullish/bearish/neutral)
2. Key support and resistance levels
3. Volume analysis
4. Trading recommendations"""

2. Financial Statement Review

prompt = """Review this income statement and:
1. Calculate key financial ratios
2. Compare YoY performance
3. Identify areas of concern
4. Highlight positive indicators"""

3. Document OCR and Extraction

prompt = """Extract all numerical data from this financial document
and organize it in a structured format."""

4. Investment Analysis

prompt = """Based on this quarterly report:
1. Summarize the company's financial health
2. Calculate growth metrics
3. Provide an investment thesis"""

Performance Comparison

Metric	8-bit (This)	BF16
Memory Usage	~4GB	~8GB
Inference Speed	Faster	Baseline
Quality	Very Good	Highest
Recommended For	Limited RAM / Speed	Maximum Quality

Hardware Requirements

Apple Silicon	Performance
M1 (8GB)	Good
M1 Pro/Max (16GB+)	Very Good
M2/M2 Pro/Max	Excellent
M3/M3 Pro/Max	Excellent
M4/M4 Pro/Max	Best

Minimum: 8GB unified memory Recommended: 16GB+ for larger batch sizes

Model Variants

Variant	Precision	Size	Speed	Quality
bf16	BFloat16	~8GB	Baseline	Highest
8bit	8-bit Quantized	~4GB	Faster	Very Good

When to Use This Model

Choose the 8-bit version if:

• You have limited RAM (8-16GB)
• You need faster inference
• You're running on battery power
• You want to run multiple models simultaneously

Choose the bf16 version if:

• You have 16GB+ RAM
• You need maximum accuracy
• You're doing fine-tuning or evaluation
• Quality is more important than speed

Training Data

This model was fine-tuned on specialized financial datasets:

• FinTrain: Comprehensive financial training data
• MultiFinBen-EnglishOCR: Multi-lingual financial document OCR
• SecureFinAI Contest: Financial security and analysis tasks
• ChartQA: Chart question answering
• NuminaMath-CoT: Mathematical reasoning with chain-of-thought
• FinCoT: Financial chain-of-thought reasoning
• COTA-LLaVA: Complex reasoning visual data

Limitations

• Primarily trained on English financial data
• Slight quality reduction compared to bf16 due to quantization
• May require prompt engineering for optimal results
• Performance varies with image quality
• Not a replacement for professional financial advice

Citation

@misc{amsi-fin-o1-mlx-8bit,
  title={Amsi-fin-o1 MLX 8-bit: Quantized Financial Vision-Language Model for Apple Silicon},
  author={AITRADER},
  year={2025},
  url={https://huggingface.co/AITRADER/Amsi-fin-o1-MLX-8bit}
}

Acknowledgments

• Original model: AITRADER/Amsi-fin-o1
• Base architecture: Qwen3-VL
• Conversion: mlx-vlm v0.3.9
• Framework: Apple MLX

License

This model is released under the Apache 2.0 License.

---

Optimized & Quantized for Apple Silicon with MLX
Made with Apple MLX Framework