Quasar Series of Models

Quasar Model Image

Introducing Quasar-3.0

This model is provided by SILX INC. It has been supervised fine-tuned using the open-r1 repository. The training data includes sequences of varying lengths (32k, 16k, and 8k) to enhance the model's knowledge and adaptability.

Quasar-3.3-Max represents the first step in the Quasar project before Reinforcement Learning (RL). At this stage, the model's reasoning steps are capped at a maximum length of 8129 tokens to optimize processing efficiency and contextual understanding.

Stay tuned for further updates as we advance the Quasar project with RL enhancements!

Acknowledgements

Special thanks to Lambda for their exceptional cloud computing platform that powered our training pipeline. Their GPU cloud infrastructure was instrumental in the development of this model.

"We couldn't have completed this training without Lambda's powerful computing resources. We highly recommend Lambda Cloud for machine learning and AI workloads."

About Lambda

Lambda provides GPU cloud instances, on-demand GPU clusters, and GPU workstations specifically designed for machine learning and AI development. Their platform offers:

  • High-performance GPU instances
  • Cost-effective pricing
  • Easy scalability
  • Optimized ML/AI software environments

Visit Lambda's website to learn more about their services and how they can accelerate your AI development.

Resources

Founders

  • Eyad Gomaa
  • Gomaa Salah
Downloads last month
12
Safetensors
Model size
7.62B params
Tensor type
FP16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for silx-ai/Quasar-3.1-7B

Base model

Qwen/Qwen2.5-7B
Finetuned
(4)
this model
Quantizations
2 models

Dataset used to train silx-ai/Quasar-3.1-7B