Soket AI logo

Soket AI models

Soket AI is a Bengaluru research lab building the Project EKA foundation models for math, code and reasoning in Indian and Global South languages, along with Indic language and speech models.

Bengaluru, Indiasoket.ai

Releases
4
Open weights
2
Latest
Nov 03, 2026

November 2026

1 release
  1. LLMProprietary

    EKA-1

    EKA-1 is the first model from Soket AI's Project EKA, a 26B-parameter sparse hybrid model trained from scratch for math, coding, reasoning and agentic work across Indian, Global South and programming languages.

    Modalities
    text → text
    Context
    64K context
    Parameters
    26B
    At launch
    Enterprise
    • Soket's RALSNet design stacks 52 layers that mix attention, recurrent and expert layers
    • Trained on 7.5 trillion tokens with an in-house tokenizer
    • Supports tool calling and Soket's open-source Loop agent framework
    • At launch Soket listed a 30.49 GB checkpoint on IndiaAI's AIKosh under Apache 2.0 with restricted visibility and asked developers to apply for early access; it also offered on-premises deployment through its inference platform
    Source: EKA-1-26B

January 2026

1 release
  1. LLMOpen weights

    Sarthi-Agri-v1

    Sarthi-Agri-v1 is a research-preview model from Soket AI, fine-tuned from Google's Gemma 3 27B to generate structured crop advisories for farmers in Hindi and English.

    Modalities
    text → text
    License
    Apache-2.0
    At launch
    Open weights
    • Trained with LoRA supervised fine-tuning on Soket's Sarathi-AgriData dataset
    • Uses control tokens to separate internal reasoning from the final advisory
    • Takes structured inputs such as crop, month, soil type, region and weather
    • Soket noted the model does not replace certified agricultural experts
    Source: soketlabs/sarthi-agri-v1 · Hugging Face

November 2025

1 release
  1. SpeechProprietary

    Dhrith

    Dhrith is Soket AI's speech recognition model for Hindi-English code-mixed speech that adds tags for emotion and tone to its transcripts.

    Modalities
    audio → text
    • Soket reported 11.19% WER on its Hindi-English benchmark, second only to Gemini 2.5 Flash among the systems it tested
    • Marks expressive cues such as tone and emotion inline with the transcript
    • Soket said expression tagging degraded in noisy audio
    • Soket said public APIs and developer access would follow
    Source: Introducing Dhrith: Emotionally Intelligent ASR

April 2024

1 release
  1. Small modelOpen weights

    Pragna-1B

    Pragna-1B is Soket AI's open 1.25B-parameter decoder-only model for Hindi, Gujarati, Bangla and English, designed for on-device use.

    Modalities
    text → text
    Context
    2K context
    Parameters
    1.25B
    License
    Apache-2.0
    At launch
    Open weights
    • Architecture follows TinyLlama, with 22 layers, grouped-query attention and a 69,632-token vocabulary
    • Custom tokenizer that Soket reported at 2.85 tokens per Kannada word versus 5.81 for Gemma-7B
    • Continued pretraining on about 150 billion tokens focused on Hindi, Bangla and Gujarati
    • Base weights went to Hugging Face, while an instruction-tuned version ran as a research preview chat
    Source: Introducing Pragna-1B: Soket AI Labs' Multilingual LLM

Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.

Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

Citation

The AI Engineering Editors. "Soket AI models: release history and timeline." The AI Engineering, updated Nov 03, 2026. https://theaiengineering.com/releases/soket-ai

HTML link

<a href="https://theaiengineering.com/releases/soket-ai">Soket AI models: release history and timeline</a> (The AI Engineering)
Full dataset
1,250 releases from 73 labs
Coverage
to
Updated
License
CC BY 4.0
Publisher
The AI Engineering

Other labs