Apollo-1-4B

Apollo-1-4B is a 4 billion parameter instruction-tuned model developed by Noema Research.
It is based on Qwen3-4B and optimized for reasoning, instruction following, and lightweight deployment at scale.

This model represents the mid-size member of the Apollo series, balancing performance and efficiency for a broad range of use cases.

Model Overview

Base model: Qwen3-4B
Architecture: Decoder-only transformer
Parameters: ~4B
Context length: up to 32k tokens (inherits Qwen3 long-context support)
Domain: General-purpose reasoning and instruction following
Primary applications:
- Conversational AI
- Multi-step reasoning tasks
- Education and tutoring systems
- Knowledge assistants and prototyping agents
License: anvdl-1.0

Key Features

Instruction tuning for consistent conversational and task-oriented responses
Improved reasoning depth compared to Apollo-1-2B, enabling stronger performance on complex queries
Long-context handling, inherited from Qwen3 architecture
Multilingual coverage, retaining broad knowledge across languages
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-1-4B"

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 helpful reasoning assistant."},
    {"role":"user", "content":"Summarize the main differences between reinforcement learning and supervised learning."}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=768, 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-1-4B demonstrates stronger reasoning capabilities relative to Apollo-1-2B, with internal evaluations indicating:

Higher accuracy on step-by-step reasoning tasks
More robust instruction adherence
Reduced hallucinations in factual settings
Effective balance between performance and efficiency

A full benchmark report will be provided in a future update. For upstream performance details, see the Qwen3-4B model card.

Limitations

Reasoning scale: While improved, Apollo-1-4B cannot match larger models (14B+) on complex or open-ended tasks
Knowledge breadth: Some specialized or domain-specific knowledge remains limited
Hallucinations: May generate plausible but incorrect information
Prompt sensitivity: Outputs remain dependent on careful prompt formulation

Responsible Use

Do not rely on Apollo-1-4B 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-1-4B and the Qwen3 base model:

@misc{noema2025apollo4b,
  title={Apollo-1-4B},
  author={Noema Research},
  year={2025},
  howpublished={\url{https://huggingface.co/NoemaResearch/Apollo-1-4B}}
}

Acknowledgements

Apollo-1-4B builds upon the Qwen3 family of models. We thank the Qwen team for open-sourcing their models and enabling derivative research.

NoemaResearch
/

Apollo-1-4B