eulogik/Bharat-Tiny-LLM-fused

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Bharat-Tiny-LLM-fused by eulogik is a 1.5 billion parameter language model, based on the Qwen2.5 architecture, with a 32768 token context length. This full-precision fused model integrates a LoRA adapter into the base Qwen2.5-1.5B weights, provided in PyTorch float16. It is specifically designed for server-side inference, further fine-tuning, or for producing custom quantized builds like GGUF and MLX. The model is optimized for scenarios requiring a balance of performance and flexibility in deployment.

Loading preview...

Overview

This repository hosts the Bharat-Tiny-LLM-fused model, a 1.5 billion parameter language model developed by eulogik. It is a full-precision (PyTorch float16) version of the Bharat-Tiny-LLM, where a LoRA adapter has been merged into the base Qwen2.5-1.5B weights. This fused model is distinct from its quantized counterparts, offering maximum flexibility for developers.

Key Capabilities

  • Full-Precision Inference: Suitable for running inference on CPU/CUDA using the transformers library.
  • Fine-Tuning Base: Provides a robust foundation for further fine-tuning on specific datasets or tasks.
  • Custom Quantization: Ideal for generating custom quantized builds, such as GGUF for llama.cpp or MLX for Apple Silicon, tailored to specific deployment needs.
  • Qwen2.5 Architecture: Leverages the capabilities of the Qwen2.5 model family.

Good For

  • Server-side deployments: When computational resources allow for full-precision models.
  • Developers requiring maximum control: For those who need to fine-tune the model or create their own optimized quantized versions.
  • Experimentation: Provides a flexible base for research and development.

Important Considerations

When using this model with transformers, specific generation parameters are recommended to avoid garbled output, including temperature ≈ 0.3, repetition_penalty ≥ 1.25, and no_repeat_ngram_size = 3.