MetaNova-1-60M

<div align="center"> <img src="https://capsule-render.vercel.app/api?type=waving&color=gradient&customColorList=12,20,24&height=200&section=header&text=MetaNova%201&fontSize=60&fontColor=ffffff&animation=fadeIn&fontAlignY=35&desc=60M%20Parameter%20Language%20Model%20%E2%80%A2%20WhirlwindAI&descAlignY=58&descSize=16&descColor=d0d0ff" width="100%"/>

</div>

---

📊 Benchmarks

MetaNova-1 leads the series on ArithMark-2 — outperforming GPT-2 (124M) with less than half the parameters.

<div align="center"> </div>

---

💬 Chat Format

MetaNova-1 uses the DeepSeek / Qwen im_start format with optional thinking mode.

<details open> <summary><b>🧠 Thinking Mode</b> — append <code>/think</code></summary>

<|im_start|>user
{query} /think<|im_end|>
<|im_start|>assistant
<think>
{thinking_content}
</think>

{response}<|im_end|>

</details> <details> <summary><b>⚡ Non-Thinking Mode</b> — append <code>/no_think</code></summary>

<|im_start|>user
{query} /no_think<|im_end|>
<|im_start|>assistant
<think>

</think>

{response}<|im_end|>

</details>

---

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "WhirlwindAI/MetaNova-1-60M"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "<|im_start|>user\nWhat is 14 × 27? /think<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

---

🧩 Model Details

---

<div align="center"> <img src="https://capsule-render.vercel.app/api?type=waving&color=gradient&customColorList=12,20,24&height=120&section=footer&fontSize=16&fontColor=ffffff&animation=twinkling" width="100%"/>

_{Built by <a href="https://huggingface.co/WhirlwindAI">WhirlwindAI</a> · Apache 2.0}

</div>