Dhīr 2B
Dense transformer
- Parameters
- 2.03B
- Context
- 16,384 tokens
- 4-bit file
- 1.21 GB
- 16-bit weights
- 4.06 GB
DhīrDhīr 2B and Dhīr 5B MoE
Compact foundation models pretrained from scratch in India, for offline use on laptops, embedded boards and edge devices, in English and ten Indian languages.
Licence: Apache 2.0
The models
Dhīr 5B MoE
Mixture of experts, 2 of 12 experts per token
- Parameters
- 5.09B total, 1.79B active per token
- Context
- 16,384 tokens
- 4-bit file
- 2.95 GB
- 16-bit weights
- 10.19 GB
Both models share a tokenizer, block design and training recipe, and are released as instruction-tuned checkpoints.
At a glance
- Pretraining tokens per model
- 6T+
- Languages: English and ten Indian languages
- 11
- Token context
- 16,384
- No network connection needed
- Offline
Languages: Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia, alongside English.
Built from scratch in India
These are the first billion-parameter generative language models built in India and pretrained from scratch on TPUs.
From random initialisation
Pretrained from scratch, not adapted from another developer’s checkpoint.
On TPUs in India
Every training stage ran on Google Cloud TPU v6e in Mumbai, with data and checkpoints held in the same region.
Compute in production
2,048 TPU v6e chips and 64 NVIDIA H100 GPUs in production.
Own tokenizer
A 131k-entry tokenizer built in-house for English and ten Indian languages.
Verifiable lineage
Every checkpoint from the first training step to release is identified by a SHA-256 digest.
Open-source evaluation
The tools we use to measure Dhīr and its tokenizer are public, so every reported number can be checked.
brahmai-evalbrahmai-tokenizer-eval
Results
| Model | Indic | Agentic | Code and maths | English | Overall |
|---|---|---|---|---|---|
| Dhīr 2B | 53.7 | 85.9 | 50.8 | 55.1 | 61.4 |
| Dhīr 5B MoE | 51.2 | 87.2 | 61.2 | 58.9 | 64.6 |
- Dhīr 2B is first in its class on 13 of 15 Indic benchmarks.
- Dhīr 2B’s Indic average matches or exceeds every 4B-class model evaluated.
- Dhīr 5B MoE leads Dhīr 2B on agentic, code and maths, and English averages; Dhīr 2B leads on Indian languages.
Section averages in percent, higher is better, across 41 benchmarks. Both models were compared with 24 open models of 1–7 billion parameters under one protocol; the full tables are in the technical report. Technical report
Runs on your hardware
| Device | Dhīr 2B | Dhīr 5B MoE |
|---|---|---|
| Laptop integrated graphics (Intel Iris Xe) | 18.6 | 20.5 |
| NVIDIA Jetson Orin NX | 34.1 | 39.9 |
| Raspberry Pi 4 (CPU) | 2.3 | 2.8 |
Decode speed in tokens per second, single stream, 4-bit weights, measured with llama.cpp.
Where it runs
Industrial sites
Assistants and structured requests on the shop floor, without sending data off the premises.
Network-denied and strategic settings
Models that keep working when the network does not.
Public-service kiosks
Local answers in the languages people speak.
Private conversations
Inference that stays on the user’s own hardware.
Safety and limitations
On a multilingual safety benchmark covering English and nine Indian languages, Dhīr 5B MoE had the lowest rate of harmful responses of the eight models evaluated, and Dhīr 2B the third lowest.
- Recall of facts without a supporting source is the weakest area. Use retrieval from authoritative sources and human review for law, finance, agriculture, medicine or history.
- Larger models lead on mathematics and code. Neither model has a reasoning mode, and we make no claim of scientific or engineering reasoning ability.
- Text only: no image, audio or speech input or output. Instruction following has been measured in English only.
What’s next
Dhīr 9B
A language and scientific foundation model, built on the same training stack. In development; public release subject to evaluation.
Infrastructure
The same infrastructure prepares training data for our next models. On Google Kubernetes Engine, our synthetic-data generation pipeline’s throughput rose 4.07×, from 1.47 to 5.97 billion generated tokens a day per 1,000 accelerators.
Deploy with ZenteiQ
- 1
Pilot
Validate on a defined use case.
- 2
Adapt
Fine-tune on approved enterprise data.
- 3
Deploy
Cloud, private cloud, on-premises or offline edge.
- 4
Operate
Integration, monitoring, updates and support.
Available through enterprise licensing and pilot programmes. How enterprises engage
Bring Dhīr into your systems
Model access, fine-tuning on your data and deployment on your hardware.