HFOpenHathi
Made in IndiaSarvam AI's open-weight Hindi LLM trained on curated Indic web data. Designed as a community base model for Hindi NLP experimentation.
Best for
- •Hindi NLP researchers needing a strong Indic base model for fine-tuning.
Limitations
- •Base model — requires fine-tuning for most production tasks.
- •Context window is limited to 4K tokens.
- •Not instruction-tuned out of the box.
Use carefully when
- •You need an instruction-following or chat-ready model out of the box.
Quickstart
- Clone the OpenHathi repo and download weights from Hugging Face.
- Use the included tokenizer optimised for Hindi and Devanagari script.
- Fine-tune on your Hindi dataset using LoRA for efficient adaptation.
Setup checklist
- • API key required: No
- • SDK quality: medium
- • Self-host difficulty: medium
Usage Notes
- • Validate model behavior on your own benchmark slices before rollout.
- • Pin version/provider routes for reproducible outputs.
- • Add logging + fallback routes for high-volume workloads.
Pricing (INR)
Input / 1M
₹0.00
Output / 1M
₹0.00
Monthly
-
Capabilities
- contextTokens4096
- openWeightYes
- indicLanguages1
Benchmarks
indic N L U
69.8
hindi Perplexity
12.4
tokenizer Efficiency
85.2
Community reviews
0 reviews • avg —
No reviews yet.
Samples
textHindi text completion
भारत एक विविध → (completion continues in Hindi).
Compliance
- License: apache-2.0
- Commercial use: allowed
Provenance
- Last verified: 2/3/2026
- Source: https://www.sarvam.ai