eulogik/Bharat-Tiny-LLM-v2-MLX

🇮🇳 Bharat-Tiny-LLM v2 — MLX

1.5B Hindi-capable Open-Weights LLM with Brahmi Token Injection. Optimized for Apple Silicon (Mac, iPhone, iPad).

---

---

✨ Key Innovation: Brahmi Token Injection

Bharat-Tiny-LLM v2 introduces 300 surgically-injected Devanagari subword tokens into the Qwen2.5-1.5B tokenizer. This achieves 33.8% fewer tokens when encoding Hindi text — resulting in faster inference, lower memory footprint, and higher throughput on edge devices.

---

📊 Performance Benchmarks & Technical Advantages

Metric	Base Qwen2.5-1.5B	Bharat-Tiny-LLM v2 MLX	Technical Advantage
Tokens for 1,000 Hindi Chars	~950 tokens	~630 tokens	33.8% Token Reduction (up to 58% on chat prompts)
Model Size (Q4 Quantized)	880 MB	880 MB	Fits in <3.8 GB RAM
Inference Speed (Hindi)	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)

---

🚀 Quick Start with MLX

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

# Load quantized open weights + LoRA adapter
model, tokenizer = load(
    "eulogik/Bharat-Tiny-LLM-v2-MLX",
    adapter_path="eulogik/Bharat-Tiny-LLM-v2-MLX/lora_adapter"
)

# Generate Hindi text
sampler = make_sampler(temp=0.3)
response = generate(
    model, 
    tokenizer,
    prompt="मैं आपको बताना चाहता हूँ कि",
    max_tokens=100,
    sampler=sampler
)
print(response)

---

💻 Hardware & Platform Support

Device	Runtime Engine	Peak RAM	Status
Apple Silicon Mac (M1/M2/M3/M4)	MLX / Python	3.8 GB	✅ Real-time (68 tok/s)
iPhone / iPad	MLX Swift	3.8 GB	✅ Real-time (On-Device)
PyTorch / Linux / CUDA GPUs	Transformers	Var	Use eulogik/Bharat-Tiny-LLM-v2

---

📜 License & Citation

Licensed under Apache 2.0. Free for commercial and research use.

@misc{kishore2026brahmi,
    title={Brahmi: Efficient Devanagari Token Injection for Multilingual LLMs},
    author={Gautam Kishore},
    year={2026},
    publisher={eulogik},
    howpublished={\url{https://huggingface.co/eulogik/Bharat-Tiny-LLM-v2-MLX}}
}