Dhīr 2B
A compact foundation model, designed and pretrained from scratch in India, with an in-house tokenizer built for Indian languages.
ZenteiQ · AIKosh · BrahmAI family · Apache 2.0
Overview
Dhīr 2B is the lightweight model in ZenteiQ's BrahmAI family. It was designed and pretrained from scratch by ZenteiQ's team in India, rather than adapted from an existing foundation model. It understands English and ten Indian languages, and is small enough to run close to the user: on a laptop, a workstation, an edge device, or inside an enterprise boundary where data must stay private.
| Developer | ZenteiQ AiTech Innovations Pvt. Ltd., Bengaluru, India |
| Model type | Dense decoder-only Transformer, instruction-tuned (SFT and GRPO reinforcement learning) |
| Parameters | 2.03B |
| Languages | English, Hindi, Gujarati, Tamil, Bengali, Kannada, Malayalam, Marathi, Odia, Punjabi, Telugu |
| Context length | 16,384 tokens |
| Licence | Apache 2.0 |
| Support | IndiaAI Mission, MeitY, under the Foundation Model initiative |
| Companion model | Dhīr 5B MoE (more capacity for complex enterprise workloads) |
Dhīr 2B is optimised for Indic document QA and lightweight, on-device deployment; Dhīr 5B MoE adds stronger reasoning, code and tool calling at ~1.8B active-parameter cost.
Why Dhīr 2B
- Built in India, end to end. Model design, tokenizer, data pipeline, pretraining and post-training were carried out by ZenteiQ's team in India.
- Best Indian-language model in its class. Best Indic ↔ English translation, cross-lingual summarisation, sentiment, causal reasoning, natural-language inference and reading comprehension among open models up to 3.2B.
- Best document QA in its class. Best question answering over supplied Indian-language passages (IndicQA, XQuAD-IN, XorQA-IN) among open models up to 3.2B, the core skill of RAG applications.
- Native tool and function calling. Calls tools and APIs in a native tool format (BFCL v3 accuracy 72.4), so it can power assistants that act on documents, apps and services.
- Private by design. 3.8 GB in BF16 and 1.1 GB at 4-bit, so it can run on-device or on-premises without a large cloud model.
Architecture
| Component | Configuration |
|---|---|
| Layers | 28 |
| Hidden dimension | 2048 |
| Attention | Grouped-Query Attention: 16 query heads, 8 key/value heads, head dim 128 |
| Feed-forward | Gated FFN, intermediate dimension 8192 |
| Positional encoding | Rotary Position Embeddings (RoPE) |
| Normalisation | RMSNorm |
| Attention kernel | Flash Attention |
| Vocabulary | 131,072 (in-house tokenizer for Indian languages) |
| Context | 16k tokens |
| Precision | bfloat16 |
Training stack: MaxText (JAX) on Google TPU v6e. Checkpoints are trained in Orbax format and converted to Hugging Face format for release.
Quick start
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "zenteiq-ai/Dhir-2B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
messages = [{"role": "user", "content": "भारत की राजधानी के बारे में दो वाक्य लिखिए।"}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
output = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))
trust_remote_code=True is required. Dhīr ships its architecture as custom code (modeling_dhir.py and configuration_dhir.py, included in this repository), so Transformers must be allowed to run it when loading the model. The same applies to other tools that load the model through Transformers: for example, pass --trust-remote-code to vllm serve.
Intended use
| Use | Notes |
|---|---|
| Indic ↔ English translation | Drafting and everyday communication. Human review is recommended for published or legal content. |
| Question answering over supplied documents (RAG reader) | Answers from a passage you provide, in English and Indian languages. |
| Indian-language text understanding | Sentiment, reading comprehension and summarisation. |
| Base model for Indic fine-tuning | A strong starting point for organisations building their own Indian-language model. |
| Basic Python assistance | Short, self-contained functions. |
| Tool and function calling | Native function calling for assistants that call tools and APIs. |
| Private, offline and edge AI | Personal assistants and privacy-sensitive applications on local hardware. |
Evaluation
All figures are for the Dhīr 2B release checkpoint, from ZenteiQ's internal benchmark suite (2 November 2026). Ranks are against open models of up to 3.2B parameters, run on the same benchmark and harness: 11 instruction-tuned models for generative tasks and 12 models (base and instruct) for log-likelihood tasks, Dhīr 2B included. Ranks exclude the other Dhīr model. Leads of less than 1 point are reported as joint first: differences that small are within run-to-run and evaluation-framework variation.
| Capability | Benchmark | Dhīr 2B | Next best | Rank |
|---|---|---|---|---|
| Translation, 24 Indic ↔ English directions (chrF) | IndicGenBench FLORES | 50.8 | 33.4 | #1 of 11 |
| Cross-lingual summarisation, 13 languages | IndicGenBench CrossSum | 19.0 | 15.2 | #1 of 11 |
| Document QA, 11 languages (token F1) | IndicQA | 60.7 | 37.6 | #1 of 11 |
| Document QA, 13 languages (token F1) | IndicGenBench XQuAD-IN | 60.3 | 41.8 | #1 of 11 |
| Cross-lingual QA, 13 languages (token F1) | IndicGenBench XorQA-IN | 34.3 | 20.5 | #1 of 11 |
| Sentiment, 14 languages (accuracy) | IndicSentiment | 95.9 | 88.7 | #1 of 12 |
| Causal reasoning, 19 languages (accuracy) | IndicCOPA | 64.1 | 60.3 | #1 of 12 |
| Natural-language inference (accuracy) | XNLI | 44.6 | 42.9 | #1 of 12 |
| Reading comprehension, 20 languages (accuracy) | Belebele | 53.6 | 49.5 | #1 of 12 |
| Python (pass@1) | MBPP + HumanEval | 45.1 | 44.4 | Joint #1 of 12 |
| Instruction following, English (strict) | IFEval | 70.4 | 70.4 | Joint #1 of 11 |
| Grade-school maths, English | GSM8K | 61.3 | Basic maths reasoning | |
| Function calling, native tool format (accuracy) | BFCL v3, AST checker | 72.4 |
Reference: larger models. Dhīr 2B also scores above every 4B-class model in the comparison (Gemma-3-4B-it, Qwen3-4B, Qwen3-4B-2507, Phi-4-mini, Phi-3-mini, Phi-3.5-mini) on translation, cross-lingual summarisation, XorQA-IN, sentiment and IndicCOPA, and above Gemma-3-4B-it on code (43.4).
Document QA is measured when the answer is in a supplied passage. Function calling is scored in the model's native tool format (BFCL v3, AST checker), averaged over 15 categories.
Deployment. Weights take 3.78 GB (BF16), 2.01 GB (8-bit) and 1.13 GB (4-bit). Deployment checks passed on NVIDIA L4 and RTX Pro 6000 GPUs, on CPU with 8-bit and 4-bit quantisation, and at 4K, 8K and 16K context. At 128 concurrent users the model sustains about 2,945 output tokens/s in aggregate, around 27 tokens/s per user.
Evaluation notes
- Results come from ZenteiQ's internal evaluation suite on the release checkpoint. Independent reproduction is part of the release process, and the full methodology will appear in the Dhīr Technical Report.
- Every model in a comparison was scored with the same harness, prompt and metric.
- Each claim is tied to the benchmark that supports it. We do not extend a result beyond what the benchmark measures.
Precautions
Dhīr 2B is a compact model, and we recommend the following for responsible deployment:
- Expert validation. Legal, medical and financial uses should be reviewed by a qualified professional. The model is not intended to give such advice on its own.
- Grounding. Answers are most reliable when the relevant text is supplied in the prompt, as in RAG applications.
- Language coverage. Supported languages are English and the ten Indian languages listed above.
- Reasoning mode. Thinking mode is not supported in this release.
- Human oversight. Like all language models, Dhīr 2B can produce incorrect or biased output. Keep a human in the loop for consequential decisions.
Safety
Dhīr 2B has been evaluated for safety behaviour, including probing with the NVIDIA Garak vulnerability scanner across jailbreak, prompt-injection and harmful-request categories. Detailed results will be published with the technical report. Deployers should add application-level safeguards suited to their use case.
Training
- Pretraining: from scratch on Google TPU v6e using MaxText (JAX).
- Tokenizer: in-house, designed for Indian languages (vocabulary 131,072), shared with Dhīr 5B MoE.
- Post-training: supervised fine-tuning followed by GRPO reinforcement learning.
Data composition, training compute and methodology will be provided in the Dhīr Technical Report.
Availability
Released under the Apache 2.0 licence through AIKosh, Hugging Face and ZenteiQ.
Acknowledgement
ZenteiQ acknowledges the support of the IndiaAI Mission, Ministry of Electronics and Information Technology (MeitY), under the Foundation Model initiative.
Citation
@misc{zenteiq2026dhir,
title = {Dh\={\i}r : Sovereign Compact Foundation Models },
author = {{ZenteiQ AiTech Innovations Pvt. Ltd.}},
year = {2026},
howpublished = {Technical report forthcoming}
}
Contact
ZenteiQ | India