Image credit: Exceptional Technology — AI-generated illustrative scene for ZenteiQ’s Dhīr models, a software platform with no official product photograph available.
Most AI models arrive in India as imports — trained elsewhere, tuned elsewhere, and run on somebody else’s cloud. On 3 November 2026, Bengaluru-based ZenteiQ showed a different path: two foundation models built from scratch in India, trained on chips in Mumbai, and small enough to run entirely offline on a laptop. The company calls them Dhīr 2B and Dhīr 5B MoE, and the company says they are the first billion-parameter generative language models built in India and pretrained from scratch on TPUs.
The launch took place in New Delhi under the IndiaAI Mission, in the presence of MeitY Secretary S Krishnan and senior IndiaAI officials, alongside Soket Labs’ EKA-1. For ZenteiQ, born from scientific-computing research at IISc Bengaluru, it is the moment a research idea becomes infrastructure Indian engineers can actually download, fine-tune and run on their own hardware.
Two compact models, trained from nothing but data
Dhīr 2B is a dense transformer with 2.03 billion parameters, and Dhīr 5B MoE is a mixture-of-experts model with 5.09 billion parameters in total, of which 1.79 billion are active for any given token — two of twelve experts are called on per token. Both share a 16,384-token context, a tokenizer built in-house with 131,000 entries, and a training recipe the company ran from random initialisation rather than adapting another developer’s checkpoint. Each model was pretrained on more than six trillion tokens.
Every training stage ran on Google Cloud TPU v6e chips in Mumbai, with data and checkpoints held in the same region, and the company says 2,048 TPU v6e chips and 64 NVIDIA H100 GPUs are in production. That geography matters: the models, the data pipeline and the checkpoints all live in India, and every checkpoint from the first training step to release is identified by a SHA-256 digest, so the lineage of each released model can be checked.
More Read
Small enough for a laptop, fluent in eleven languages
The most striking number is not the parameter count but the file size. Dhīr 2B ships as a 1.21 GB file in 4-bit precision, and Dhīr 5B MoE as a 2.95 GB file — small enough to carry on a laptop and run with no network connection at all. The company says both models run offline on laptops and edge devices, and it publishes decode speeds measured with llama.cpp: 18.6 and 20.5 tokens per second on a laptop’s integrated graphics, 34.1 and 39.9 on an NVIDIA Jetson Orin NX, and 2.3 and 2.8 even on a Raspberry Pi 4.
Both models work in English and ten Indian languages — Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia — and the company says Dhīr 2B leads its class on 13 of 15 Indic benchmarks, with an Indic average that matches or exceeds every 4-billion-parameter-class model it evaluated. Across 41 benchmarks against 24 open models of one to seven billion parameters, the company reports overall scores of 61.4 for Dhīr 2B and 64.6 for Dhīr 5B MoE, with the larger model ahead on agentic tasks, code and maths, and English, and the smaller model ahead on Indian languages.
Built for factory floors, not just chat windows
ZenteiQ describes Dhīr as physics-native: foundation models grounded in physics and engineered for real-world industry, rather than general chatbots dressed up for enterprise. The company says its models are already in pilot engagements across automotive, renewable energy, semiconductor, advanced manufacturing, aerospace, technology services and commerce — places where an assistant on the shop floor, a simulation surrogate in R&D, or a kiosk answering in the local language cannot wait for a cloud round-trip, or cannot send data off the premises at all.
That is why the deployment story is as important as the model itself. Enterprises can fine-tune Dhīr on their own data and deploy it on the cloud, on a private cloud, on-premises, or fully offline — and the company’s products, including NokkX, ScireX and AhamX, are built to put the models into engineering workflows, with its BrahmX system preparing the scientific and engineering data that feeds them.
Open weights, open evaluation
The company says it is releasing the model weights under the Apache 2.0 licence, along with technical documentation and evaluation tools, through AIKosh — so any Indian team can download, inspect, fine-tune and build on Dhīr without asking permission or paying a licence fee. The evaluation tools themselves are public too, so every reported number can be checked rather than taken on trust.
ZenteiQ was founded by Sashikumaar Ganesan, a professor of Computational and Data Sciences at IISc Bengaluru who moved to the company full-time in 2026, and the company says more than 65 engineers and scientists now work on its models and enterprise systems. Selected under the IndiaAI Mission’s Foundation Model initiative, it is building a family that already stretches from Dhīr 2B to a Dhīr 9B language and scientific model now in development.
For India’s engineers, the promise is beautifully practical: world-class AI that speaks their languages, runs on the machine already on their desk, and keeps their data exactly where it belongs — with them.
Keep reading: The Gate Before Send: Inside Defmate Secure, Gurugram’s Shield for AI · Eyes That Never Blink: Inside Oculosense’s Drishti Vision Live AI Glasses · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
