Daedalus-1-2B

Daedalus-1-2B is a 2 billion parameter code reasoning model developed by Noema Research. It is based on DeepCoder-1.5B-Preview and optimized for advanced code generation, debugging, and algorithmic reasoning.

This model represents the entry-level member of the Daedalus series, balancing performance and efficiency for a broad range of software engineering tasks.

---

Model Overview

• Base model: DeepCoder-1.5B-Preview

• Architecture: Decoder-only transformer

• Parameters: ~2B

• Context length: up to 64k tokens

• Domain: Code reasoning and generation

• Primary applications:

  • Code completion and synthesis
  • Debugging and error detection
  • Algorithm design and explanation
  • Educational tools and coding assistants

• License: MIT

---

Key Features

• Instruction tuning for reliable multi-step reasoning and task completion
• Extended context handling, supporting up to 64k tokens
• Multilingual support, including Python, C++, Java, JavaScript, Go, Rust, and more
• Reinforcement learning optimization for improved code generation accuracy
• Efficient deployment, suitable for both cloud and edge environments

---

Usage

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

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "NoemaResearch/Daedalus-1-2B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

prompt = "Write a Python function to merge two sorted lists into a single sorted list."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, top_p=0.9)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended settings:

• temperature=0.5–0.8
• top_p=0.9–0.95
• Lower temperatures yield more deterministic and concise code completions

---

Evaluation

Daedalus-1-2B demonstrates strong performance in code reasoning tasks, with internal evaluations indicating:

• High accuracy on code completion and synthesis tasks
• Robust debugging capabilities, identifying and suggesting fixes for common errors
• Effective handling of complex algorithmic problems

A full benchmark report will be provided in a future update. For upstream performance details, see the DeepCoder-1.5B-Preview model card.

---

Limitations

• Reasoning scale: While effective for many tasks, Daedalus-1-2B may not match larger models (e.g., 4B+) on highly complex or open-ended coding problems
• Knowledge breadth: Some specialized or domain-specific knowledge may be limited
• Hallucinations: May generate plausible but incorrect code or explanations
• Prompt sensitivity: Outputs remain dependent on careful prompt formulation

---

Responsible Use

• Do not rely on Daedalus-1-2B for critical software development tasks without human oversight
• Verify all generated code before deploying in production environments
• 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 Daedalus-1-2B and the DeepCoder base model:

@misc{noema2025daedalus2b,
  title={Daedalus-1-2B},
  author={Noema Research},
  year={2025},
  howpublished={\url{https://huggingface.co/NoemaResearch/Daedalus-1-2B}}
}

---

Acknowledgements

Daedalus-1-2B builds upon the DeepCoder family of models. We thank the Agentica team for open-sourcing their models and enabling derivative research.

---