Dhir-2B-8bit

Dhīr-2B-8bit is the 8-bit quantized release of Dhīr 2B , ZenteiQ's 2.03-billion-parameter foundation model pretrained from scratch in India. Distributed in GGUF Q8_0 format , this 2.2 GB model enables private, offline inference using llama.cpp on compatible laptops, mobile devices and edge hardware. It supports English and ten Indian languages, with a context length of 16,384 tokens. Quantization fidelity Quantized directly from the original BF16 checkpoint without additional training, Dhīr-2B-8bit was evaluated against the reference model using WikiText-2 perplexity, next-token prediction agreement, KL divergence and standard benchmarks. Key evaluation results: Within 0.1% perplexity difference compared with the BF16 model on WikiText-2. 98% next-token prediction agreement with BF16 over approximately 1.5 million evaluation tokens. Evaluation coverage includes English, ten Indian languages, code and mathematics. These results demonstrate strong preservation of next-token prediction behavior following 8-bit quantization. They do not imply that every generated response is identical to the BF16 model. Model file: Dhir-2B-Q8_0.gguf (2.2 GB) License: Apache 2.0 Compatibility: Requires the ZenteiQ Dhīr-support branch of llama.cpp until support is integrated upstream. More information: https://zenteiq.ai/dhir/