Fara-TARS-7B: The Hybrid Reasoning & GUI Agent

Fara-TARS-7B is a state-of-the-art merged model that combines the high-level reasoning and planning capabilities of Microsoft Fara-7B with the precise GUI grounding and agentic capabilities of ByteDance UI-TARS-7B.

This model achieves a Hybrid Mode: it can seamlessly switch between writing complex text plans (Reasoning) and executing precise coordinate actions (Agentic Tool Calls) based on the user prompt.

Key Capabilities

How to Use (Inference Code)

To unlock the full potential of this model (Agent Mode vs Text Mode), you must use the specific generation configuration below. This handles the tool schema injection and prevents repetition loops.

Installation

pip install torch transformers pillow

Python Inference Class

Use this class to interact with the model. It handles the system prompt injection and JSON parsing automatically.

import torch
from transformers import AutoModelForVision2Seq, AutoTokenizer, GenerationConfig
from PIL import Image
import json
import re

class FaraAgent:
    def __init__(self, model_path, device="auto"):
        print(f"Loading Fara-TARS from {model_path}...")
        self.model = AutoModelForVision2Seq.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            device_map=device,
            trust_remote_code=True,
            low_cpu_mem_usage=True
        )
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        
        # Safety fix for padding
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id

        # Define Agent Tools
        self.tools_schema = [
            {"name": "left_click", "description": "Click coordinate [x, y]", "parameters": {"type": "object", "properties": {"point": {"type": "array"}}}},
            {"name": "type_text", "description": "Type text", "parameters": {"type": "object", "properties": {"text": {"type": "string"}}}},
            {"name": "scroll", "description": "Scroll screen", "parameters": {"type": "object", "properties": {"pixels": {"type": "integer"}}}},
            {"name": "terminate", "description": "Task done", "parameters": {"type": "object", "properties": {"status": {"type": "string"}}}}
        ]

    def _format_prompt(self, user_prompt):
        # Injects the schema and strict English/Format instructions
        tools_json = json.dumps(self.tools_schema, indent=2)
        system = (
            f"You are Fara-TARS, a GUI automation agent.\n"
            f"AVAILABLE TOOLS:\n{tools_json}\n\n"
            "INSTRUCTIONS:\n"
            "1. Reason first, then act.\n"
            "2. Output valid JSON inside <tool_call> tags.\n"
            "3. Format: <tool_call>{{\"name\": \"left_click\", ...}}</tool_call>"
        )
        return (
            f"<|im_start|>system\n{system}<|im_end|>\n"
            f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
            f"<|im_start|>assistant\n"
        )

    def _repair_json(self, json_str):
        # Auto-fixes common LLM JSON errors (smart quotes, missing keys)
        json_str = json_str.replace("“", '"').replace("”", '"').replace("'", '"')
        json_str = re.sub(r'(\w+)"\s*:', r'"\1":', json_str) 
        return json_str

    def run(self, prompt, image_path=None):
        formatted_prompt = self._format_prompt(prompt)
        
        # Handle Image Input (Optional)
        if image_path:
            image = Image.open(image_path).convert("RGB")
            inputs = self.model.build_conversation_input_ids(
                tokenizer=self.tokenizer, query=formatted_prompt, image=image
            )
        else:
            inputs = self.tokenizer(formatted_prompt, return_tensors="pt").to(self.model.device)

        # Critical: Stop generation at tool close to prevent loops
        stop_strings = ["</tool_call>", "<|im_end|>"]
        
        # Optimized Config
        config = GenerationConfig(
            max_new_tokens=2048,
            do_sample=True,
            temperature=0.4, 
            top_p=0.95,
            repetition_penalty=1.15, # Prevents "United.com" loops
            no_repeat_ngram_size=0,  # Must be 0 to allow JSON keys
            pad_token_id=self.tokenizer.pad_token_id,
            eos_token_id=self.tokenizer.eos_token_id
        )

        with torch.no_grad():
            output = self.model.generate(
                **inputs,
                generation_config=config,
                tokenizer=self.tokenizer,
                stop_strings=stop_strings
            )

        input_len = inputs['input_ids'].shape[1]
        raw_response = self.tokenizer.decode(output[0][input_len:], skip_special_tokens=True)
        
        # Parse Output
        tool_action = None
        text_content = raw_response
        
        if "<tool_call>" in raw_response:
            parts = raw_response.split("<tool_call>")
            text_content = parts[0].strip()
            tool_str = parts[1].split("</tool_call>")[0].strip()
            try:
                tool_action = json.loads(self._repair_json(tool_str))
            except:
                tool_action = {"error": "malformed_json", "raw": tool_str}

        return {"thought": text_content, "action": tool_action}

# Usage
agent = FaraAgent("your-username/Fara-TARS-7B")
result = agent.run("Click the Submit button at (1200, 800)")
print(result)

Benchmark Performance

The model was evaluated on a comprehensive suite covering Web Automation, GUI Grounding, and Complex Reasoning.

Merge Details

This model was merged using Mergekit.

Configuration

models:
  - model: microsoft/Fara-7B
  - model: ByteDance-Seed/UI-TARS-1.5-7B
merge_method: slerp
base_model: microsoft/Fara-7B
dtype: bfloat16
parameters:
  t:
    # 5-point gradient:
    # 0.1 (Start): Mostly Fara -> Ensures input understanding and English grammar.
    # 0.3 -> 0.5 (Middle): Blends TARS capability for reasoning and logic.
    # 0.1 (End): Mostly Fara -> Ensures the output stops correctly and doesn't loop.
    - value: [0.1, 0.3, 0.5, 0.3, 0.1]

(Note: While slerp was used, specific inference parameters (temp=0.4, rep_penalty=1.15) are required to stabilize the output, as documented in the Usage section).

Limitations

1. Strict Prompting: The model expects the specific System Prompt defined in the usage class. Without it, it may hallucinate tool names.
2. Repetition: In extremely long lists (100+ items), the model may repeat. The recommended repetition_penalty=1.15 fixes this for 99% of cases.

License

Apache 2.0