Solo Developer Releases Arcle V1, a 5.8-Billion-Parameter Multimodal Model

Abhinav Abinav, a solo developer from Bihar, India, has released Arcle V1, an integrated multimodal foundation model with 5.8 billion parameters, under open weights. The model processes text, images, documents, speech, and audio, supporting native image generation at 512x512 resolution and natural language speech generation at 24kHz.

Architecture and Performance

Arcle V1 is composed of a single neural network module, rather than connecting multiple models via API. All input data is mapped into a common 2,560-dimensional semantic latent space and processed through a single set of weights. The model supports a context window of 2 million tokens and handles more than 18 languages, including Hindi and India-specific knowledge.

Key benchmark performance is as follows:

• ARC-Easy: 80.0%
• GSM8K: 77.5%
• MATH-500: 74.2%
• HellaSwag: 67.0%
• Document Content Word Recall: 94.6%

Development Background

The developer faced past allegations of fraud and attacks from malicious websites but ignored them to independently train and test the model to completion. For Arcle V2, additional features such as full integration of the Omni Graph are planned.