VAJRAM · INDIC LANGUAGE MODEL
LangBrains AI Symbol
LANGBRAINS AI
Reasoning Intelligence · Native Language

An AI that speaks India's languages.

Vajram is our Indic large language model. It reasons, translates and answers in Kannada, Hindi, Telugu, Tamil, Malayalam, Marathi, Bengali and English, including the mixed way people actually write.

Try VajramSee the products
VAJRAM

A language model built for India

Vajram is our own Indic large language model, trained on seven Indian languages and English, with Kannada first.

  • ಕನ್ನಡKannada
  • हिन्दीHindi
  • తెలుగుTelugu
  • தமிழ்Tamil
  • മലയാളംMalayalam
  • मराठीMarathi
  • বাংলাBengali
  • English

Indian languages, natively

Answers in the language you asked in, including the mixed, romanized way people actually type, not a translation of an English answer.

Reasoning

Follows instructions, works through multi-step questions and keeps context across a conversation.

Translation

Moves meaning between English and Indian languages while keeping names, numbers and technical terms exact.

Structured output and tools

Returns typed, machine-readable answers and calls tools, so it can sit inside real software rather than only a chat box.

Under the hood

6.8×

Indic partitioned tokenizer

Each Indian script gets its own sub-vocabulary while English, code and maths stay intact. Encodes Kannada at 4.8 chars/token — using 6.8× fewer tokens than Qwen3 and 2.2× fewer than Gemma-3 or Sarvam.

View on GitHub
90.2%

296-case Indic suite lead

Scored 267/296 (90.2%) on our rigorous 27-scenario evaluation across 7 Indian languages — ranking #1 overall ahead of Sarvam-M 24B (239), Krutrim-2 12B (231), Gemma-3 12B (229) and Qwen3-8B (214).

1.5s

Fast median turn latency

Dense token representation yields a 1.5s median end-to-end response time on local GPUs, and 127 ms median latency for 1-token typed decision intents with 91.3% intent accuracy.

100%

Precision safety & guard

Achieves 100% precision on harmful prompts (15/15) and safe healthcare & legal queries (13/13) with zero over-refusal on benign look-alikes.

Benchmark results: where Vajram leads

Evaluated against leading open Indic models on identical hardware (A100 80 GB, llama.cpp b11393, Q8_0 GGUF, greedy decoding).

Overall 296-Case Indic Benchmark

Comprehensive evaluation across 27 scenarios covering Kannada, Kanglish, Hindi, Telugu, Tamil, Malayalam, Marathi, Bengali & English on identical A100 GPU hardware.

Vajram wins overall with 90.2% accuracy across all 296 real-world test cases.
Overall 296-Case Indic Benchmark comparison table
  • Vajram7B#1 Winner
    267 / 296 (90.2%)
  • Sarvam-M24B
    239 / 296 (80.7%)
  • Krutrim-212B
    231 / 296 (78.0%)
  • Gemma-312B
    229 / 296 (77.4%)
  • Qwen38B
    214 / 296 (72.3%)
  • Sarvam-30B30B MoE
    213 / 296 (72.0%)
  • Navarasa-2.07B
    129 / 296 (43.6%)
  • Ambari7B
    85 / 296 (28.7%)
90.2%Overall 296-Case Suite Score (267/296)
100%Safety & Refusal Precision (15/15)
6.8×Token Density Advantage vs Qwen3 (Kannada)
1.5sMedian End-to-End Native Response Latency

* Measured on identical hardware: A100 80 GB GPU, llama.cpp b11393 with Q8_0 GGUF quantization, greedy decoding, up to 1,500 output tokens per prompt using each model's native chat template. Token density evaluated on FLORES+ devtest (1,012 sentences per language). Built-in tool and server cases evaluated through Vajram's demo server runtime.

Try Vajram
PRODUCT 01

Translation API

Send English text and get natural Kannada back from our own Vachana translation model.

English in. Natural Kannada out.

  • 1,000 free tokens every month
  • From ₹2 per 1,000 output tokens
  • REST API and a no-code playground
Get an API keySee pricing
ENGood morning, my friend.
KNಶುಭೋದಯ, ನನ್ನ ಸ್ನೇಹಿತ.
PRODUCT 02

SQX

Structured Query eXtraction. Send a call transcript and your QA criteria; get YES, NO or NA for each, with a reason that points to the moment in the call.

Your QA checklist, answered for every call.

  • About one second per call
  • Hinglish and romanized Indian languages
  • Typically ₹0.03–₹0.06 per call
Get an API keySee pricing
  1. AgentHello, good evening. Main Rohan baat kar raha hoon EduPath se.
  2. CustomerHaan ji, boliye.
  3. AgentAapne PhD admission ke liye enquiry ki thi. Aapki Masters kis saal mein complete hui?
  4. Customer2021 mein, M.Sc Chemistry.
  5. AgentTheek hai. Fees aur universities ki details hamari team aadhe ghante mein call karke batayegi.
  6. CustomerTheek hai, thank you.
  • YESDid the agent introduce themselves and the company?Agent gave name and company in [00:00].
  • NODid the agent state the fee or a fee range?Agent stated fees would be shared later, not in this call ([00:14]).
  • NAIf the customer raised a concern, did the agent address it?Customer raised no concern in the transcript.
A real response. Point at an answer to see the line it cites.

Measured on real calls

Median per call, end to end
1.2 s
Per question
Under 1 paisa
Calls per ₹100 top-up
1,600–3,300
Median time to answer a checklist of 8 questions · 3.7× faster
  • SQX1.2 s
  • General-purpose LLM, one reply4.4 s

12 real call transcripts, 8 criteria each, measured September 2026. Slowest call: 4.3 s. Cost is the typical charge at current prices; SQX bills by tokens, so long calls cost more.

PRICING

Pay for what you use

Both APIs draw on one prepaid credit balance, charged per token.

Plans for the English → Kannada translation API. Charged per token, and a free monthly grant covers your first 1,000 tokens on any plan.

Vajram KN-1 chat model — coming soon

Free

₹0always free

1,000 free tokens / month

Evaluate the translation API, no card required.

  • English → Kannada translation
  • Allowance resets every month
  • No card required
  • Standard request queue
  • Email support
Create account or sign in

Standard

MOST POPULAR

₹2per 1,000 output tokens

₹100 buys 50,000 tokens · max ₹100

Input ₹0.10 per 1,000 tokens

Production translation at moderate volume.

  • English → Kannada translation
  • Usage-based billing: charged only for tokens processed
  • Requests stop at the cap, so there is no surprise overage
  • Standard request queue
  • Email support
Create account or sign in

Pro

₹5per 1,000 output tokens

₹2,500 buys 500,000 tokens · max ₹2,500

Input ₹0.10 per 1,000 tokens

Priority capacity for high-volume, latency-sensitive traffic.

  • Everything in Standard
  • Priority queue with higher concurrency and throughput
  • Built for latency-sensitive, high-volume workloads
  • Direct support channel
Create account or sign in
  • A plan sets your translation rates and rate limit. Its price is added to your credit balance, which every API draws on.
  • Chat with Vajram KN-1 is not included, and will be priced separately when it launches.
  • Prices are exclusive of GST, which will be added once billing is live.
  • Credit carries over month to month. The 1,000-token free grant resets monthly and is always spent first.
FAQ

A few useful answers

Vajram is LangBrains AI's own Indic large language model. It is trained on Kannada (its primary language), Hindi, Telugu, Tamil, Malayalam, Marathi, Bengali and English, including the mixed, romanized way people type, and handles reasoning, translation, structured output and tool use.

support@langbrains.in