Two different bets in one release

EKA-1 and the Dhīr models are not three sizes of the same idea. EKA-1 is the scale bet: 26.3 billion parameters and roughly 7.5 trillion training tokens, aimed at reasoning, code and agent work across India’s official languages. The Dhīr pair is the deployment bet: 2 and 5 billion parameters, designed to run on-premise and offline, on GPUs and TPUs. The PIB release itself quotes ZenteiQ CEO Prof. Sashikumar saying the lightweight models can run on mobile phones without cloud connectivity, keeping data on the device. That is a company executive’s deployment claim reproduced in the government’s announcement, not an independently demonstrated performance result. The release supplies no named phone, memory requirement or measured inference speed. For government and defense users working in low-connectivity or sensitive settings, that second property is the actual product. A model that can operate offline could solve a procurement problem that a larger cloud-hosted model does not. But the release provides no technical documentation demonstrating the deployment architecture or confirming that all components can operate without external connectivity.

Both bets were paid for the same way. The release says subsidized national compute and support for ancillary research expenses under the IndiaAI Mission enabled the development of all three. That is the mission’s model working as designed: public infrastructure in, domestic models out. What the release does not say is how much support each team received, or what either model costs to run.

The numbers that are not there

Read the official text closely and the gaps form a pattern. There is no benchmark score for EKA-1: no coding result, no reasoning result, no per-language evaluation across the 22 languages it is said to span. The release gives no benchmark score for either Dhīr model either, and no measurement behind the offline claim, such as a named phone, a memory footprint, or a tokens-per-second figure. ZenteiQ’s own AIKosh page does post average scores across 41 benchmarks — self-reported averages, with the full tables in a technical report the page lists as coming soon and no methodology published yet. The PIB release names no software license for any of the three models. It calls Dhīr open-source and EKA-1 open, but does not specify the permissions attached to their weights. ANI’s syndicated reporting separately identifies Apache 2.0 for Dhīr; that attribution should not be extended to EKA-1. The PIB text itself says only that platforms such as AIKosh “can be made accessible” to researchers and the public, which describes a route rather than confirming a listing anyone can download today. The listings have since appeared on AIKosh, but all three are marked restricted — downloading requires contributor approval. The Dhīr weights are there, with an Apache 2.0 license file; EKA-1’s listing holds settings files only, no weights.

The language count shows why the primary text matters. ANI’s version, carried by several outlets, says EKA-1 supports more than 100 Indian and international languages, alongside programming languages. The PIB release makes a narrower statement: its capabilities span all 22 scheduled Indian languages alongside English. These are different claims. The wider figure does not appear in the government document, and the two should not be combined or presented as equivalent. A model genuinely strong across 22 Indian languages would already be a substantial result; it does not need a borrowed hundred.

What would settle it

Three publications would turn this launch into a release outsiders can evaluate: a technical report or model card with benchmark scores, including per-language results for EKA-1; published score tables behind the averages ZenteiQ has posted; and AIKosh listings a researcher outside the program can actually pull — today all three are restricted, and EKA-1’s holds no weights. Until those exist, the honest description of November 3 is narrower than the ceremony suggested. India has funded and announced three domestic models with a credible deployment idea behind the small ones. Whether they are good models is a question the launch documents do not answer.

Sources

  1. [1] Press Information Bureau, “IndiaAI Mission Released Indigenous Foundation Models EKA-1 by Soket and Dhīr 2B & Dhīr 5B by ZenteiQ Under the IndiaAI Innovation Centre Pillar”, November 3, 2026, release ID 2332315 (primary source; read directly on pib.gov.in at Gate 1 — EKA-1 at 26.3 billion parameters, trained on approximately 7.5 trillion tokens, spanning all 22 scheduled Indian languages alongside English, described as an open, sparse hybrid model for math, code, reasoning and agentic use; Dhīr 2B and Dhīr 5B as open-source models supporting English and 10 Indian languages, for secure on-premise, offline-capable deployment across defense, government, industrial and educational sectors; the passage quoting ZenteiQ CEO Prof. Sashikumar saying the lightweight models can run on mobile phones without cloud connectivity, keeping data on their devices; MeitY Secretary S. Krishnan’s statement that through platforms such as AIKosh these models “can be made accessible to researchers, students, innovators and the wider public”; subsidized national compute and support for ancillary research expenses under the IndiaAI Innovation Centre (Foundation Models) pillar; the Union Cabinet’s approval of the mission on March 7, 2024; and no benchmark score and no named license anywhere in the release)Read source
  2. [2] Transclew, “Government calls EKA-1 AI model open, but you can’t download it yet”, November 4, 2026 (first-hand read of the AIKosh model pages and file lists, read on November 4, 2026 — ZenteiQ’s page posts its own average scores across 41 benchmarks, with the full tables in a technical report it lists as “coming soon”, while EKA-1’s page offers only “internal evaluations”, with results “to follow in a technical report”; AIKosh marks all three models “Restricted”, and its FAQ says a restricted item needs explicit approval from the contributor before it can be downloaded; the Dhīr weights are present in full — 3.79 GB for Dhīr 2B and 9.50 GB for Dhīr 5B, the file list including weight files and a license file, both Apache 2.0; EKA-1’s only uploaded version holds three settings files, 103.28 KB in all, and no weights, and Soket’s own website lists the model as “coming soon” and asks people to join a waitlist)Read source
  3. [3] The Hindu BusinessLine, “IndiaAI Mission releases indigenous foundation models EKA-1, Dhīr 2B and Dhīr 5B”, November 4, 2026Read source
  4. [4] ANI via NewKerala, “MeitY unveils EKA-1, Dhīr 2B and Dhīr 5B indigenous AI models under IndiaAI Mission”, November 3, 2026 (source of the “more than 100 languages”, Apache 2.0 and Google Cloud TPU details; treated as secondary reporting, not adopted where PIB differs)Read source
  5. [5] Gadgets 360, “MeitY Launches Three New Sovereign AI Models Under IndiaAI Mission”, November 4, 2026 (AIKosha availability wording; 550 million SFT samples figure, not in PIB text)Read source