--- title: "IndiaAI’s 26bn-parameter EKA-1 is ‘open’, but access is by waitlist" summary: "What the IndiaAI Mission released on 3 November (Soket’s EKA-1, ZenteiQ’s Dhīr 2B and 5B), what a new user can get on 5 November, and the 12-model plan behind them." url: https://www.thetechbag.com/blog/indiaai-eka-1-open-model-waitlist format: Note desk: "AI in the enterprise" region: India published: 2026-11-05 facts_checked: 2026-11-05 author: TechBag (The TechBag Journal) tags: ["IndiaAI Mission","Soket AI","EKA-1","ZenteiQ","Dhīr","AIKosh","MeitY","open weights","Apache 2.0","foundation models","sovereign AI","Indian languages","BharatGen","Sarvam"] cover: https://www.thetechbag.com/blog/covers/indiaai-eka-1-open-model-waitlist.png --- # IndiaAI’s 26bn-parameter EKA-1 is ‘open’, but access is by waitlist *What the IndiaAI Mission released on 3 November (Soket’s EKA-1, ZenteiQ’s Dhīr 2B and 5B), what a new user can get on 5 November, and the 12-model plan behind them.* Note · AI in the enterprise · By TechBag · Published 2026-11-05 · Facts checked 2026-11-05 · Canonical: https://www.thetechbag.com/blog/indiaai-eka-1-open-model-waitlist EKA-1, the 26.3-billion-parameter model the IndiaAI Mission released in New Delhi on 3 November, carries the Apache 2.0 licence on its government listing, and the same listing is marked “Restricted”.[1][2] Its maker, Soket AI, offers it through a waitlist: “Our team reads each request and approves access.”[3] The Mission released ZenteiQ’s Dhīr 2B and Dhīr 5B the same day; their listings are Apache 2.0 and “Restricted” too.[1][5] **Figure: EKA-1's licence is open; getting the model runs through registration, a waitlist and an approval** *EKA-1 as announced and as offered to a new user, 3 to 5 November 2026* Announced on 3 November - 26.3bn parameters, trained on about 7.5trn tokens; an “open, sparse hybrid language model” - Apache 2.0 licence on AIKosh, the Mission’s model platform - Languages: 22 scheduled Indian languages and English (PIB); “more than 100” (AIKosh); “60+” (Soket) - For maths, code, reasoning and agents, with a 64K-token context Available to a new user on 5 November - Listing marked “Restricted”; previewing its 30.49 GB of files needs registration - A waitlist on Soket’s site, each request read and approved by its team - The form asks consent for Soket to use “my prompts and the responses” to improve its models - No benchmarks: results “to follow in a technical report” Source: PIB; AIKosh EKA-1-26B listing; Soket AI early-access and Project EKA pages · as of 5 November 2026 ## What an open licence promises, and what it does not Apache 2.0 grants anyone who has the work a “perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable” licence to copy, modify and distribute it, commercial use included.[8] It governs what you may do with the weights once you hold them; it does not oblige the publisher to hand them to everyone. Open-weight models are usually posted for anyone to download, as Soket did with its 1.25bn-parameter Pragna-1B on Hugging Face in March 2024, under the same licence.[7] For EKA-1, the licence settles the terms and the waitlist settles who gets in. Dhīr shows the other half. ZenteiQ publishes section averages across 41 benchmarks against 24 open models, and its evaluation code, but says the models are “available through enterprise licensing and pilot programmes”.[6] Dhīr 2B has 2.03bn parameters; Dhīr 5B is a mixture of experts with 5.09bn, 1.79bn active per token; both cover English and ten Indian languages and were trained on Google Cloud TPUs in Mumbai.[6] ## Where EKA-1 sits in the Mission's plan The Cabinet approved the IndiaAI Mission on 7 March 2024 with an outlay of ₹10,371.92 crore.[9] In March 2026 MeitY told the Lok Sabha it backs 12 organisations to build sovereign foundation models. Sarvam, BharatGen and Gnani had launched theirs at the IndiaAI Impact Summit 2026, and the resulting models “are expected to contribute to the open-source ecosystem by making them available through AIKosh”.[10] Soket’s commitment there is “India’s first open-source 120 billion parameter foundation model”, with support of ₹177.08 crore, ₹162.47 crore of it compute. ZenteiQ’s is BrahmAI, 8bn to 80bn parameters, with ₹206.49 crore.[10] EKA-1 is “the first release from Project EKA”; Soket says a 120bn-plus model is in training on 1,536 H100 GPUs from the Mission.[2][4] TechBag countThe 12 allocations, added up The annexure to MeitY’s reply lists support for all 12: ₹2,194.29 crore in total, ₹2,035.15 crore of it compute, about 21% of the Mission’s outlay. BharatGen’s ₹1,058.52 crore is almost half.[10] **What to watch:** the EKA-1 technical report and its benchmarks, whether the AIKosh listing opens to registered users without approval, and the 120bn model. How we did this Figures as of 5 November 2026, 7pm IST Release facts are PIB’s; access terms are the AIKosh listings and the makers’ own pages, read on 5 November. The three language counts come from three sources and are not reconciled. On 5 November we found neither model on Hugging Face under its maker’s name. Allocations are MeitY’s table; the totals and the 21% are our arithmetic. Spotted an error? Write to us; corrections are published at the foot of the piece. Sources 1. [PIB (MeitY), IndiaAI Mission released indigenous foundation models EKA-1 by Soket and Dhīr 2B and Dhīr 5B by ZenteiQ](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2332315&reg=48&lang=2) — 3 November 2026 2. [AIKosh, EKA-1-26B model listing](https://aikosh.indiaai.gov.in/home/models/details/eka_1_26b_1.html) — updated 2 November 2026 3. [Soket AI, Early access to EKA-1](https://soket.ai/early-access) — undated page 4. [Soket AI, Project EKA](https://soket.ai/project-eka) — undated page 5. [AIKosh, Dhir-5B and Dhir-2B model listings](https://aikosh.indiaai.gov.in/home/models/details/dhir_5b_1.html) — updated 2 November and 30 October 2026 6. [ZenteiQ, Dhīr 2B and Dhīr 5B MoE](https://zenteiq.ai/dhir/) — undated page 7. [Hugging Face, soketlabs/pragna-1b](https://huggingface.co/soketlabs/pragna-1b) — published 3 March 2024 8. [Apache Software Foundation, Apache License, Version 2.0, section 2](https://www.apache.org/licenses/LICENSE-2.0) — January 2004 9. [PIB, Cabinet approves ambitious IndiaAI Mission to strengthen the AI innovation ecosystem](https://pib.gov.in/PressReleasePage.aspx?PRID=2012355) — 7 March 2024 10. [MeitY, Lok Sabha unstarred question 5417, Small Language Models, with Annexure I (copy hosted by MediaNama)](https://www.medianama.com/wp-content/uploads/2026/04/loksabhaquestions_annex_187_AU5417_haoHcv.pdf) — 25 March 2026 **Disclosure.** TechBag is a software reseller: it earns a margin when a reader buys through it, including from vendors named here. The analysis is TechBag’s own; no vendor paid for or saw this piece before publication. Facts are labelled reported, TechBag count or our view. **Read this next:** [US$300bn pledged to India's data centres; subsidised GPUs reportedly 20% used](https://www.thetechbag.com/blog/india-ai-compute-pledges-gpu-use) — The Mission’s compute side: subsidised GPUs onboarded under IndiaAI, reportedly 20% used, beside US$300bn of data-centre pledges. Go deeper on TechBag - [Google Cloud](https://www.thetechbag.com/google/google-cloud) — the TPUs in Mumbai on which ZenteiQ says Dhīr was trained - [Sarvam’s Prism decision model](https://www.thetechbag.com/blog/sarvam-prism-decision-model-4-rupees-million-tokens) — another IndiaAI-backed lab’s Apache 2.0 release the same week, with an API price - [Chinese open-weight models on OpenRouter](https://www.thetechbag.com/blog/chinese-open-weight-models-gateway-traffic) — how open-weight models are licensed and used worldwide