---
library_name: transformers
pipeline_tag: text-generation
license: apache-2.0
language:
- en
- zh
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
tags:
- text-generation-inference
- Code
- Math
- RL
- R1
---
![KV.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/7WWuZljYRluVp5gi3--9a.png)

# **Primus-Optima-QwenKV-1.54B**

> **Primus-Optima-QwenKV-1.54B** is an **experimental chain-of-thought reasoning and code generation model**, built by combining the strengths of two sources:
>
> - **DeepSeek R1 (distilled 1.5B)** for strong math and coding reasoning.
> - **Qwen2.5-0.5B**, fine-tuned with **Process Reward Models (PRM)** to boost structured step-by-step outputs in math and logic.

This hybrid design results in a **bilingual, high-precision model** with enhanced **reasoning depth**, **multi-step clarity**, and **lightweight adaptability** for math and code applications.

## **Key Features**

1. **Chain-of-Thought Reasoning for Math + Code**  
   Designed to produce human-like intermediate steps in both math and programming problems — useful for education, tutoring, and technical assistants.

2. **Hybrid Architecture (Reasoning + Reward-Guided Fine-Tuning)**  
   Combines **DeepSeek R1’s** distilled capabilities with **Qwen2.5-0.5B**'s reward-optimized reasoning for structured, goal-driven outputs.

3. **Multilingual Capabilities (English + 中文)**  
   Fluent and accurate in both English and Simplified Chinese, making it suitable for diverse learning and development environments.

4. **Coder Experimental Mode**  
   Able to solve algorithmic tasks, complete functions, and offer code walkthroughs using the same step-by-step format as it does for math.

5. **Lightweight Yet Capable (1.54B)**  
   With just 1.54B parameters, it is efficient for local deployments while offering surprisingly strong performance on STEM and programming tasks.

## **Quickstart with Transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Primus-Optima-QwenKV-1.54B"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Write a Python function to compute factorial using recursion."
messages = [
    {"role": "system", "content": "You are an expert tutor in math and programming, explaining step-by-step."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```

## **Intended Use**

- **Math & Programming Tutors**: Assist students with logic-driven step-by-step explanations.
- **Bilingual STEM Apps**: Ideal for dual-language math or coding environments.
- **Competitive Reasoning Tools**: Suited for reasoning-intensive tasks like Olympiad prep, technical quizzes, and programming challenges.
- **On-Device LLMs**: Lightweight enough for web or embedded applications needing real-time reasoning.

## **Limitations**

1. **Experimental Nature**:  
   This is a hybrid research model; performance may vary across general or creative domains.

2. **Size Constraints**:  
   As a 1.54B parameter model, extremely complex reasoning tasks may challenge its capabilities.

3. **Bias & Generalization**:  
   Inherits biases from both DeepSeek R1 and Qwen2.5. Use caution in high-stakes or sensitive applications.

4. **Prompt Engineering Required**:  
   Structured prompts with clear questions yield the best results, especially for multi-step problems.