Back home

Dhīr

x+web 0.86
Visit source

Intro

Dhīr is ZenteiQ’s compact India-pretrained foundation-model family (zenteiq.ai/dhir/): Dhīr 2B (dense 2.03B) and Dhīr 5B MoE (5.09B total, 1.79B active per token, 2-of-12 experts). Shared tokenizer and recipe; instruction-tuned checkpoints; 16,384 context; aimed at offline use on laptops, boards, and edge devices.

ZenteiQ says both were pretrained from scratch in India (not continued from another vendor’s checkpoint) on Google Cloud TPU v6e in Mumbai with data/checkpoints in-region; 6T+ tokens each; languages: English plus Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, and Odia. Tied to an IndiaAI Mission unveil on 2026-11-04. Note: the same window also mentioned early-access EKA-1; this page tracks the openly specified Dhīr family.

Explain

Problem

Multilingual Indic + edge/air-gapped settings need small models that run locally with a verifiable training lineage. Dhīr offers dense 2B and small MoE 5B for offline deploy, with public eval tooling and SHA-256 checkpoint-lineage claims.

Differentiation

  • From-scratch India pretraining narrative for billion-parameter generative models (per vendor), not overseas-base adaptation.
  • Two sizes: 2B leads 13/15 Indic benches in-class (vendor tables); 5B MoE leads on agentic, code+math, and English averages.
  • Edge numbers: public llama.cpp 4-bit decode rates on Iris Xe / Jetson Orin NX / Pi 4.

Limits

  • Vendor flags weak unsupported fact recall, no reasoning mode, text-only; law/finance/medicine need retrieval + human review.
  • Instruction-following measurements disclosed mainly in English so far.
  • Enterprise paths are pilot/license; distinct from early-access-only EKA-1.

Usage

Who it is for

Teams needing English+Indic offline/edge LLMs, or tracking IndiaAI domestic bases.

Getting started

  1. Read zenteiq.ai/dhir/ and the tech report for sizes, languages, and safety tables.
  2. Fetch checkpoints via the site paths; run 4-bit with llama.cpp on target hardware.
  3. For enterprise: ZenteiQ pilot → adapt → cloud / private cloud / on-prem / fully offline.

Advertisement