Qwen3.5-2B-ReMix-Final is a precision-engineered, native Float16 (F16) fine-tune of Qwen/Qwen3.5-2B. While the previous ReMix iterations focused on the broad integration of large-scale distillation datasets, the Final version is the result of a specialized Supervised Fine-Tuning (SFT) strategy designed to maximize stability and logical coherence.

This model is specifically tuned to eliminate the "reasoning loops" common in small models. By shifting the training focus to strict instruction-following and adversarial logic handling, ReMix-Final acts as a robust, logic-first assistant for local execution.

🚀 Key Improvements & Comparison

This model marks a significant departure from the base and the initial ReMix:

Superior Logic Handling: While the base model is prone to repetitive cycles under stress, ReMix-Final demonstrates a vastly improved ability to traverse complex constraints and converge on an answer.
Instruction Following: Leveraging SFT datasets, the model adheres more strictly to formatting requirements and complex multi-step instructions.
Impossible-Question Awareness: Trained on custom hard-distillation sets, the model has been taught to recognize logical contradictions, allowing it to "break" a loop by identifying a problem as unsolvable.

🌟 Model Details

Base Model: Qwen/Qwen3.5-2B (Pre-integrated with multi-source distillation)
SFT Foundations: * allenai/Dolci-Instruct-SFT
nvidia/Nemotron-SFT-Instruction-Following-Chat-v2
Reasoning Enhancement: * Jackrong/DeepSeek-V4-Distill-8000x
5x Custom high-difficulty distillation sets targeting logical "dead-ends."
Format: Native F16 Merged Weights.
License: Apache-2.0

🎛️ Recommended Generation Parameters

Parameter	Value	Purpose
Temperature	`0.4 - 1.0`	Essential for keeping reasoning deterministic and focused.
Repetition Penalty	`1.15 - 1.2`	Acts as a safety net to help the model break out of residual loops.
Top K / Top P	`30 / 0.9`	Provides the model with enough vocabulary depth for technical tasks.
enable_thinking	`True`	Recommended to leverage the internal reasoning architecture.
context_length & max_token	> 4096	Allow the model to freely reason through. This model usually take more than 4000 - 5000 tokens to reason.

⚠️ Limitations & Fallback Behavior

Residual Looping: Despite the additional SFT training aimed at stability, 2B models can still fall back into looping patterns when faced with extreme ambiguity or recursive paradoxes. This is a significant improvement over the base and earlier ReMix, but remains a known characteristic of compact architectures.
Specialized Logic: The model is optimized for procedural reasoning (math, logic, code). For creative writing or general conversation, the "minimized reasoning" training may result in shorter, more direct responses than expected.
Verification: Always verify mathematical and technical outputs.

📦 Usage (Transformers)

python

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
                }
            },
            {
                "type": "text",
                "text": "Where is this?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="ertghiu256/Qwen3.5-2b-ReMix-final",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.9,
    repeat_penalty=1.2,
    extra_body={
        "top_k": 30,
    }, 
)
print("Chat response:", chat_response)

Uploaded finetuned model

Developed by: ertghiu256
License: apache-2.0
Finetuned from model : ertghiu256/Qwen3.5-2b-ReMix-Vision-GRPO

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.