How We Cut Devanagari LLM Token Costs by 33.8% via Brahmi Token Injection

Gautam Kishore — Fri, 24 Jul 2026 17:09:29 +0000

Introducing Bharat-Tiny-LLM v2: Cutting Devanagari LLM Token Costs by 33.8%

Multilingual transformer models built on English-dominant training datasets impose a severe "Token Tax" on non-Latin scripts. When processing Devanagari (Hindi) script, standard Byte-Pair Encoding (BPE) tokenizers break whole words into fragmented sub-byte sequences.

For developers deploying Indic LLMs in production, this token bloat causes:

3x Higher Memory Bandwidth Consumption
3x Slower Token-per-Second Throughput
Inflated Cloud API Costs
Premature Context Window Exhaustion

To solve this efficiency barrier, we engineered Brahmi Token Injection and released open weights for Bharat-Tiny-LLM v2.

Here is a look at the achievement, benchmarks, and how to run the open-weights model family locally.

The Breakthrough: 33.8% Token Compression

Consider the Hindi phrase:

"ज़रूरी बात है क्या करते हो"

On a standard 1.5B multilingual base model, encoding this prompt requires 26 tokens.

With Bharat-Tiny-LLM v2, the exact same phrase requires only 11 tokens — representing a 58% reduction in token count for conversational text. Across broad evaluation corpora, Bharat-Tiny-LLM v2 averages a 33.8% overall token reduction.

[Standard Tokenizer]   ज़रूरी बात है क्या करते हो  ➔ 26 Tokens 🔴
[Bharat-v2 Tokenizer]  ज़रूरी बात है क्या करते हो  ➔ 11 Tokens 🟢 (58% Savings)

Performance & Quality Benchmarks

By optimizing vocabulary efficiency, Bharat-Tiny-LLM v2 achieves significant throughput and quality gains while keeping the model size down to an ultra-compact 880 MB:

Benchmark Metric	Standard Base Model	Bharat-Tiny-LLM v2	Achievement Delta
Tokens per 1,000 Hindi Chars	~950 tokens	~630 tokens	33.8% Token Reduction
Inference Speed (Apple M-Series)	50 tok/s	68 tok/s	+36% Faster Throughput
Validation Loss (Hindi Corpus)	2.776	1.837	52.5% Loss Reduction (Perplexity: 16.1 ➔ 6.3)
Model Size (Q4 Quantized)	880 MB	880 MB	Runs in <3.8 GB Peak RAM

Prompt Compression Examples

Hindi Prompt	Standard Tokens	Bharat-v2 Tokens	Token Savings
`"ज़रूरी बात है क्या करते हो"`	26 tokens	11 tokens	58% Savings
`"नमस्ते, आप कैसे हैं?"`	15 tokens	7 tokens	53% Savings
`"भारत की राजधानी नई दिल्ली है"`	22 tokens	14 tokens	36% Savings
`"bhai aaj ka weather kaisa hai?"`	12 tokens	8 tokens	33% Savings

High-Level Innovation: Brahmi Token Injection

Standard tokenizers lack representation for high-frequency Indic subwords. Brahmi Token Injection solves this by:

Surgically Expanding Vocabulary: Ingesting script-specific subwords directly into the tokenizer dictionary (expanding vocabulary from 151,936 to 152,236 tokens).
Targeted Embedding & Attention Alignment: Aligning newly introduced token representations while fine-tuning attention projection layers to maintain fluency and semantic reasoning.

(Note: Detailed mathematical formulations, training logs, and ablation studies will be published in our upcoming research paper).

Quickstart: Run Open Weights Locally

Option A: PyTorch & Transformers (Linux / Windows / CUDA GPUs)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "eulogik/Bharat-Tiny-LLM-v2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

prompt = "भारत के प्रमुख त्यौहारों के बारे में लिखिए:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.7)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Option B: Apple Silicon MLX (Mac / iPhone / iPad)

from mlx_lm import load, generate

model, tokenizer = load("eulogik/Bharat-Tiny-LLM-v2-MLX")
response = generate(model, tokenizer, prompt="नमस्ते! आज का दिन कैसा है?", max_tokens=100)
print(response)

Open Weights & License

The model family, weights, and MLX quantizations are released today under Apache 2.0 for commercial, enterprise, and research deployment.

PyTorch Open Weights: huggingface.co/eulogik/Bharat-Tiny-LLM-v2
Apple Silicon MLX Weights: huggingface.co/eulogik/Bharat-Tiny-LLM-v2-MLX

*Built by Eulogik

KARN - Language agents speak

Gautam Kishore — Thu, 09 Apr 2026 17:12:00 +0000

Built a Programming Language for AI Agents

After watching AI agents waste 76% of their tokens on boilerplate in Python, TypeScript, and Rust, we built KARN — a programming language designed specifically for them.

The Problem

AI agents are writing code in languages designed for humans. This is like building a race car with sedan parts.

Language	Tokens/LOC	Designed For
Python	~6.8	Humans
TypeScript	~9.2	Humans
Rust	~11.5	Humans
KARN	~2.1	Agents

That's 4x more code in the same output limit. For an agent generating 100 programs per day, that's millions of tokens saved per month.

The Design

Token-Minimal Syntax

Every operator maps to exactly one intent. No overloading. No keywords.

->   Function definition
!    Emit/print
?    Propagate error
??   Fallback on error
|>   Pipe forward (sequential)
|~   Race (first wins)
&    Parallel execution
*    Map collection
%    Filter collection
..   Range

Error as Value

No exceptions. No hidden control flow. Every I/O returns Ok|Err. Agents always know exactly what path executes.

data = http.get(url)?           -- propagate error up
val  = cache.get(key)??fallback -- fallback on error

Multi-Platform from One Source

karn build app.kn --target c       → gcc → native binary
karn build app.kn --target js      → node app.js
karn build app.kn --target web     → open in browser
karn build app.kn --target python  → python app.python.py

Full Ecosystem Access

from pip numpy as np
from npm react as R
from cargo serde as serde
from sys ffmpeg as ff

The Codegen

The C runtime: tagged union Val type, dynamic arrays, hash maps, error propagation, full stdlib. The JS runtime maps every stdlib module to Node.js equivalents. The Python target emits portable Python 3.

All from the same AST. Zero rewrites.

Get Started

pip install karn-lang
npm install karn-lang

GitHub: https://github.com/karn-lang/karn
Website: https://karn-lang.dev
PyPI: https://pypi.org/project/karn-lang/
npm: https://www.npmjs.com/package/karn-lang
Agent Docs: https://github.com/karn-lang/karn/blob/main/AGENT.md
Machine Spec: https://raw.githubusercontent.com/karn-lang/karn/main/karn-spec.json
License: MIT — https://github.com/karn-lang/karn/blob/main/LICENSE

Built by the team at Eulogik. Thanks to everyone who made this possible!

DEV Community: Gautam Kishore