Apollo-1-2B

Apollo-1-2B is a 2 billion parameter instruction-tuned model developed by Noema Research.
It is based on Qwen3-1.7B and optimized for general reasoning, language understanding, and lightweight deployment.

This model is the first release in the Apollo series, intended as a foundation for scalable experimentation and real-world applications in constrained environments.

---

Model Overview

• Base model: Qwen3-1.7B
• Architecture: Decoder-only transformer
• Parameters: ~2B
• Context length: up to 32k tokens (inherits Qwen3 long-context support)
• Domain: General-purpose reasoning and instruction following
• Primary applications:
  • Conversational AI
  • Lightweight reasoning tasks
  • Education and tutoring
  • Prototype agents and assistants
• License: anvdl-1.0

---

Key Features

• Instruction tuned: More reliable responses in conversational and task-oriented settings
• Lightweight deployment: Optimized for environments with limited compute or memory resources
• Extended context: Inherits long-context capability from Qwen3 base
• Balanced outputs: Improved refusal behaviors and reduced hallucinations compared to the base model
• Multilingual ability: Retains multilingual knowledge from Qwen3 family

---

Usage

The model is available in Hugging Face Transformers format. Example:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "NoemaResearch/Apollo-1-2B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {"role":"system", "content":"You are Apollo, a reasoning assistant."},
    {"role":"user", "content":"Explain the difference between supervised and unsupervised learning."}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended settings:

• temperature=0.5–0.9
• top_p=0.85–0.95
• For structured outputs (e.g. JSON), use lower temperatures for stability

---

Evaluation

Apollo-1-2B has been evaluated internally on a range of reasoning and language tasks. Key findings:

• Improved instruction following relative to Qwen3-1.7B
• More concise and accurate responses in structured tasks
• Maintains multilingual performance from the base model
• Effective for lightweight assistant applications

Future work will include publishing comprehensive benchmark comparisons against other models in the 1–3B parameter range.

---

Limitations

• Reasoning depth: As a 2B parameter model, Apollo cannot match larger-scale LLMs on complex reasoning tasks
• Knowledge coverage: May lack depth in specialized or low-resource domains
• Hallucinations: Although reduced, the model may still generate incorrect or fabricated information
• Sensitivity to prompts: Outputs vary with prompt phrasing; careful prompt design recommended

---

Responsible Use

• Do not rely on Apollo for critical decision-making without human oversight
• Generated outputs may contain inaccuracies; verification is required for factual or sensitive use cases
• Avoid providing personal, private, or sensitive information in prompts
• This model should not be used to generate disallowed, unsafe, or harmful content

---

Model Variants

• Full precision (safetensors) — research and full-fidelity inference
• bf16 / fp16 — optimized for inference on GPUs/TPUs
• Quantized versions (int8 / int4) — for deployment in constrained hardware environments

---

Citation

If you use this model, please cite both Apollo-1-2B and the Qwen3 base model:

@misc{noema2025apollo,
  title={Apollo-1-2B},
  author={Noema Research},
  year={2025},
  howpublished={\url{https://huggingface.co/NoemaResearch/Apollo-1-2B}}
}

---

Acknowledgements

Apollo-1-2B builds upon the Qwen3 series of models. We thank the Qwen team for making their work openly available under permissive terms, which enabled this derivative research.

---