FV FinVector
MLX LoRA · 4B Model files ↗
FINANCIAL REASONING · STRUCTURED OUTPUT

Markets are noisy.
Reasoning should not be.

An MLX-native LoRA for Qwen3.5-4B, specialized in macro interpretation, financial calculation, tool routing, causal transmission, and structured multi-asset scenarios.

View adapter on Hugging Face See frozen evaluation
Exact base revision 851bf6e…8cd0a · Adapter hash 8defb300…e02a5
INTERFACE PREVIEW

Three financial reasoning workflows

This $0 static Space previews the model's output contracts. It does not run inference on Hugging Face's Linux servers; use the verified MLX adapter locally on Apple Silicon.

Input

The browser evaluates only numbers and + − × ÷ parentheses.

Structured preview
NOT MODEL INFERENCE 

              
            

          

        

      


      

        

          

            
LOCKED 600-ROW BENCHMARK

            
Explicit-schema comparison

          

          
Greedy decoding, thinking off, no output repair.

        


        

          

            JSON validity99.8%Base 91.3%
          

          

            FinQA exact40.0%Base 14.7%
          

          

            Correct expression82.7%Base 48.0%
          

          

            Structural scenario agreement89.5%Base 20.1%
          

          

            Structural implication agreement87.2%Base 52.4%
          

          

            Tool selection100%Base 100%
          

        


        

          Research boundary
          

            Direction scores measure agreement with synthetic structural benchmark targets, not
            realized market returns. Scenario probabilities are structural weights and are not
            empirically calibrated forecasts.
          

        

      


      

        

          
RUN THE VERIFIED MODEL

          
Inference stays local.

          

            Download the adapter from its model repository, combine it with the pinned
            Qwen3.5-4B revision, and run it with MLX-LM on Apple Silicon.
          

        

        
from huggingface_hub import snapshot_download
from mlx_lm import load

adapter = snapshot_download("alinakhay/FinVector-Market-4B")
model, tokenizer = load(
    "Qwen/Qwen3.5-4B",
    revision="851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    adapter_path=adapter,
)