A compact 2.03B-parameter foundation model, pretrained from scratch in India, for private and edge AI in English and ten Indian languages.
Dhīr 2B is designed to provide efficient, AI capabilities that can run completely offline on edge devices, enabling language understanding and generation in English and Indian languages for privacy-sensitive, industrial, and network-restricted environments
Dhīr 2B is a 2.03-billion-parameter dense decoder-only Transformer pretrained from scratch on over six trillion tokens. It uses a custom 131K-token multilingual tokenizer and was trained entirely on TPU v6e infrastructure in India
Dhīr 2B was pretrained from random initialization on about 6.2 trillion tokens, followed by staged mid-training, long-context adaptation, supervised fine-tuning and reinforcement-learning-based post-training. Training data covers English, Indian languages, mathematics, code, reasoning, instruction following, tool use, multi-turn conversation and safety. The model was trained with on TPU v6e. The supervised fine-tuning corpus contains about 15.9 million conversations.
Dhīr 2B has been evaluated internally on Indian-language understanding, agentic and tool-use tasks, instruction following, mathematics, coding, long-context retrieval and quantized deployment. A 4-bit deployment build, using about 4.3 bits per weight, has also been evaluated against the BF16 model. Detailed benchmark, comparison and quantization results, together with evaluation settings and methodology, will be published in the forthcoming Dhīr technical report.
Dhīr 2B is a compact model and does not match much larger cloud models in closed-book factual knowledge, advanced mathematics, complex coding or highly specialised reasoning. Current or changing facts should be verified through authoritative retrieval. Deployers should apply safeguards appropriate to their use case. Tool calls with external consequences should be validated by the application. Dhīr 2B is text-only and should not be used as the sole authority for medical, legal, financial, safety-critical or other consequential decisions.
Dhīr 2B is designed for private, local and edge deployment. The BF16 model occupies about 4.06 GB (10⁹ bytes), while a 4-bit deployment build substantially reduces memory requirements. For privacy-sensitive applications, inference, conversation history, local memory and retrieval should remain on-device or within the trusted deployment boundary wherever possible. External web, document or enterprise tools should be enabled only when needed, with clear user visibility when information leaves the device. Production systems should isolate conversation contexts, validate consequential tool calls, use authoritative retrieval for current information and perform domain-specific evaluation before deployment. Mobile deployments should additionally be validated for memory, thermal behaviour, battery use, sustained throughput and long-context performance on target devices.
Apache 2.0
ZenteiQ
Large Language Models
Transformers
Restricted
Sector Agnostic
30/10/26 07:44:43
3.79 GB
12 files
Apache 2.0
© 2026 - Copyright AIKosh. All rights reserved.