EKA-1-26B

EKA-1 is a 26-billion-parameter foundation model built from scratch by Soket AI with support from IndiaAI, MeitY. Designed for mathematics, coding, reasoning and agentic workflows across 100+ languages, it offers a 64K-token context window and flexible deployment on-premises, in air-gapped environments or in private clouds.

About Model

EKA-1 is a 26-billion-parameter foundation model developed entirely from scratch by Soket AI, with support from the IndiaAI Foundation Models pillar, MeitY. The first release from Project EKA, it is designed for mathematics, coding, reasoning and complex agentic workflows across more than 100 Indian, Global South and programming languages.

Powered by RALSNet (Recursive Attentive Latent Sparse Network), EKA-1 combines 52 layers with a sparse architecture comprising attention, recurrent and expert layers. Soket AI reports more than 3x faster inference than comparably sized models. Trained on 7.5 trillion tokens using an in-house tokenizer, EKA-1 supports a 64K-token context window for long documents and multi-step tasks.

EKA-1 supports tool calling and integration with agentic harnesses, including Soket AI’s open-source Loop https://soket.ai/loop, a Rust-based framework for running sandboxed agents at scale. Organisations can deploy EKA-1 through Soket AI’s Inference Platform on their own premises, in fully air-gapped environments or in their own cloud, maintaining control over sensitive data.

Internal evaluations indicate performance on par with comparably sized models, with detailed methodology and results to follow in a technical report. Developers, researchers, startups and institutions can explore Project EKA and apply for early access at https://soket.ai.