Apollo-1-8B

Apollo-1-8B

Apollo-1-8B is a 8 billion parameter instruction-tuned model developed by Noema Research. It is based on Qwen3-8B and optimized for advanced reasoning, instruction following, and high-performance deployment.

This model represents the large-scale member of the Apollo series, balancing strong reasoning capabilities with efficiency for multi-domain applications.

---

Model Overview

• Base model: Qwen3-8B

• Architecture: Decoder-only transformer

• Parameters: ~8B

• Context length: up to 32k tokens (inherits Qwen3 long-context support)

• Domain: General-purpose reasoning, instruction following, and code generation

• Primary applications:

  • Advanced conversational AI
  • Multi-step reasoning and problem solving
  • Knowledge assistants and tutoring systems
  • Software development and code generation

• License: anvdl-1.0

---

Key Features

• Instruction tuning for reliable multi-step reasoning and task completion
• Extended reasoning depth compared to Apollo-1-4B for complex queries
• Long-context handling, inherited from Qwen3 architecture
• Multilingual coverage, supporting diverse languages and domains
• Balanced resource requirements, deployable on high-end consumer hardware and cloud GPUs

---

Usage

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

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "NoemaResearch/Apollo-1-8B"
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 differences between supervised, unsupervised, and reinforcement learning with examples."}
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended settings:

• temperature=0.4–0.8
• top_p=0.9–0.95
• Lower temperatures yield more factual and concise answers

---

Evaluation

Apollo-1-8B demonstrates stronger reasoning and instruction-following capabilities relative to Apollo-1-4B, with internal evaluations indicating:

• Higher accuracy on complex multi-step reasoning tasks
• More robust instruction adherence
• Reduced hallucinations in factual and structured outputs
• High efficiency for large-context tasks

A full benchmark report will be provided in a future update. For upstream performance details, see the Qwen3-8B model card.

---

Limitations

• Reasoning scale: While improved, Apollo-1-8B cannot match ultra-large models (14B+) on extremely complex or open-ended tasks
• Knowledge breadth: Some highly specialized or niche knowledge may be limited
• Hallucinations: May generate plausible but incorrect information
• Prompt sensitivity: Outputs remain dependent on careful prompt formulation

---

Responsible Use

• Do not rely on Apollo-1-8B for critical decisions without human oversight
• Verify outputs before applying in factual, legal, or safety-critical contexts
• Avoid providing personal or sensitive data in prompts
• The model should not be used to generate unsafe, harmful, or disallowed content

---

Model Variants

• Full precision (safetensors) — research and high-fidelity inference
• bf16 / fp16 — efficient inference on modern accelerators
• Quantized versions (int8 / int4) — deployment in resource-constrained environments

---

Citation

If you use this model, please cite both Apollo-1-8B and the Qwen3 base model:

@misc{noema2025apollo8b,
  title={Apollo-1-8B},
  author={Noema Research},
  year={2025},
  howpublished={\url{https://huggingface.co/NoemaResearch/Apollo-1-8B}}
}

---

Acknowledgements

Apollo-1-8B builds upon the Qwen3 family of models. We thank the Qwen team for open-sourcing their models and enabling derivative research.