IndiaAI EKA-1 and Dhir Models: What Launched and How to Get Access

Soket's EKA-1 and ZenteiQ's Dhīr-2B and Dhīr-5B models were released at an IndiaAI Mission event on November 3. EKA-1 has an early-access application, while Dhīr focuses on smaller, offline language models. A launch announcement should not be confused with immediate, unrestricted access to every model.

The official government record confirms the models and the release in New Delhi. Their developers' documentation provides the more useful distinction: what a team can actually evaluate and where it can run.

EKA-1 access starts with an application

Soket's EKA-1 early-access page asks users to join a waitlist. Applications are reviewed, with approval communicated by email. That is not the same thing as a generally available, self-service API.

The form also requests agreement to using prompts and responses to improve the models. Teams should read that condition before submitting proprietary examples or planning a sensitive-data pilot.

Soket describes a broader Project EKA research programme spanning reasoning, coding and multilingual models. Programme-level training ambitions are not a reliable substitute for an EKA-1 model card. A developer still needs the specific checkpoint's context limit, supported languages, pricing and API terms before making an integration decision.

Dhīr-2B and Dhīr-5B target local workloads

ZenteiQ's Dhīr product documentation describes text-only models supporting English and ten Indian languages, with a 16,384-token context window. It lists Apache 2.0 licensing and points to AIKosh for downloads.

The dense 2B model has 2.03 billion parameters and a listed four-bit weight size of 1.21 GB. The 5B mixture-of-experts model has 5.09 billion total parameters and a four-bit weight size of 2.95 GB. Those are weight sizes, not the total memory an application needs once its runtime and context cache are included.

The product page also acknowledges weak factual recall, recommends retrieval where needed and says instruction-following evaluation was conducted in English. Local execution does not remove the need to verify outputs, particularly across languages.

Why the 5B model does not use every parameter per token

In its November 3 training account, ZenteiQ explains that Dhīr-5B routes each token to two of twelve experts. Roughly 1.8 billion parameters are active per token, despite the larger total model size.

That can reduce computation compared with activating all five billion parameters for every token. It does not make the full model disappear from memory. Hardware fit and generation speed still need to be measured together.

ZenteiQ says both models were pretrained from random weights on Google Cloud TPU v6e, each seeing more than six trillion tokens. Indian-language data was included during training rather than treated solely as an adaptation added later.

For a local maintenance-manual assistant, this approach could make offline operation and language coverage useful together. It would still need retrieval over the actual manuals and an evaluation set drawn from the people using it. That is a potential application, not a demonstrated production result.

What remains to be verified

The linked AIKosh destination did not expose a usable model listing during this check, so a successful weight download has not been independently verified here. ZenteiQ's publications page also lists the full technical report as coming soon.

The public brahmai-eval repository provides evaluation tooling, including support for language-model and function-calling tests. It does not contain the model weights or their evaluation results. Its README still describes the models as unreleased, which does not match the newer launch messaging.

That discrepancy needs clarification rather than an assumption that either public downloads or complete reproducibility are already established. Published benchmark figures should remain attributed to the developer until reproduced.

The next meaningful milestone is practical access: a checkpoint, documented interface and repeatable evaluation that another team can use. A domestic model becomes useful infrastructure when developers can independently test its limits, not just read its launch announcement.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain model releases and engineering decisions in ways developers can understand and share.

Book a 30-minute call.