AITRADER/Amsi-fin-o1.5-mxfp4-MLX

Amsi-fin-o1.5 — 4-bit MLX

4-bit quantized (affine, 5.06 bits/weight) MLX conversion of AITRADER/Amsi-fin-o1.5 — a 9B parameter finance-focused Vision Language Model with reasoning and tool calling capabilities.

Model Description

Amsi-fin-o1.5 is a fine-tuned Qwen3.5-VL-9B model specialized for financial analysis. It combines:

• Vision understanding — analyze charts, financial documents, and screenshots
• Chain-of-thought reasoning — structured thinking via <think> tags
• Tool calling — native function calling for trading systems and APIs
• Long context — 262K token context window

This is the smallest variant, ideal for devices with 16 GB unified memory.

Key Capabilities

Capability	Details
Vision	Chart analysis, document OCR, screenshot understanding
Reasoning	<think> tag reasoning with extended chain-of-thought
Tool Calling	<tool_call> format for API integration
Finance	Options pricing, technical analysis, portfolio management
Context Length	262,144 tokens

Model Architecture

Component	Specification
Architecture	Qwen3.5-VL (Vision-Language)
Parameters	~9B
Text Layers	32 (mixed linear + full attention)
Text Hidden Size	4,096
Text Intermediate	12,288
Attention Heads	16 (4 KV heads)
Head Dim	256
Vision Depth	27 layers
Vision Hidden Size	1,152
Patch Size	16×16, temporal 2
Vocab Size	248,320
Quantization	4-bit affine (group_size=64)
Bits/Weight	5.06

Available Variants

Variant	Size	Bits/Weight	Link
fp16	~18 GB	16	AITRADER/Amsi-fin-o1.5-fp16-MLX
8-bit (affine)	~10 GB	8.86	AITRADER/Amsi-fin-o1.5-mxfp8-MLX
4-bit (this)	~5.6 GB	5.06	You are here

Installation

pip install mlx-vlm

Requires macOS 14+ and Apple Silicon (M1 or later).

Usage

Vision — CLI

python -m mlx_vlm.generate \
  --model AITRADER/Amsi-fin-o1.5-mxfp4-MLX \
  --image chart.png \
  --prompt "Analyze this chart. What trading signals do you see?" \
  --max-tokens 512

Vision — Python API

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

model, processor = load("AITRADER/Amsi-fin-o1.5-mxfp4-MLX")
image = load_image("chart.png")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What pattern is forming on this chart?"}
        ]
    }
]

prompt = apply_chat_template(processor, messages)
output = generate(model, processor, prompt, [image], max_tokens=512)
print(output)

Reasoning

The model supports extended reasoning via <think> tags:

messages = [
    {
        "role": "user",
        "content": "A stock is trading at $150. A call option with strike $155 "
                   "expires in 30 days. IV is 25%. Calculate the approximate "
                   "option price using Black-Scholes and explain your reasoning."
    }
]

prompt = apply_chat_template(processor, messages)
output = generate(model, processor, prompt, max_tokens=2048)
# Output will contain <think>...</think> reasoning followed by the answer

Tool Calling

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_stock_price",
            "description": "Get current stock price",
            "parameters": {
                "type": "object",
                "properties": {
                    "symbol": {"type": "string", "description": "Ticker symbol"}
                },
                "required": ["symbol"]
            }
        }
    }
]

messages = [
    {"role": "user", "content": "What's the current price of AAPL?"}
]

prompt = apply_chat_template(processor, messages, tools=tools)
output = generate(model, processor, prompt, max_tokens=256)
# Model will generate: <tool_call><function=get_stock_price>{"symbol": "AAPL"}</function></tool_call>

Streaming Generation

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

model, processor = load("AITRADER/Amsi-fin-o1.5-mxfp4-MLX")

messages = [{"role": "user", "content": "Explain covered call strategy."}]
prompt = apply_chat_template(processor, messages)

for token in stream_generate(model, processor, prompt, max_tokens=512):
    print(token, end="", flush=True)

Chat UI (Gradio)

python -m mlx_vlm.chat_ui --model AITRADER/Amsi-fin-o1.5-mxfp4-MLX

Hardware Requirements

Apple Silicon	Unified Memory	fp16	8-bit	4-bit
M1 / M2	16 GB	❌	❌	✅
M1/M2 Pro	32 GB	✅	✅	✅
M1/M2 Max	64 GB	✅	✅	✅
M1/M2 Ultra	128+ GB	✅	✅	✅
M3	24 GB	✅	✅	✅
M3 Pro	36 GB	✅	✅	✅
M3 Max	48–128 GB	✅	✅	✅
M4	24–32 GB	✅	✅	✅
M4 Pro	48 GB	✅	✅	✅
M4 Max	64–128 GB	✅	✅	✅

4-bit runs comfortably on 16 GB devices — the most accessible variant.

Conversion Details

Converted using mlx-vlm v0.3.12:

python -m mlx_vlm.convert \
  --hf-path AITRADER/Amsi-fin-o1.5 \
  --mlx-path ./Amsi-fin-o1.5-mxfp4-MLX \
  --quantize --q-bits 4

Vision tower weights merged from the Qwen3.5-VL-9B base model to ensure full vision capability.

Training Details

• Base Model: Qwen3.5-VL-9B
• Fine-tuning Focus: Financial analysis, options trading, technical analysis
• Capabilities Added: Domain-specific finance reasoning, trading tool integration

Limitations

• Primarily trained on English financial data
• Should not be used as sole basis for trading decisions
• Vision analysis works best with standard chart formats
• Long-context performance may vary with very large documents
• 4-bit quantization trades some precision for efficiency; for maximum accuracy use the fp16 variant

Citation

@misc{amsi-fin-o1.5,
  title={Amsi-fin-o1.5: Finance Vision Language Model},
  author={AITRADER},
  year={2025},
  url={https://huggingface.co/AITRADER/Amsi-fin-o1.5}
}

Acknowledgments

• Qwen Team for the Qwen3.5-VL base model
• MLX Team at Apple for the MLX framework
• mlx-vlm by Prince Canuma for VLM conversion tooling

License

Apache 2.0 — see LICENSE