Show HN: Arc Decide RAG and conversation insights under 1gb (huggingface.co)

The launch of Arc Decide's CPU-first model, arc-decide-0.4b-v0.2, marks a significant advancement in decision-making capabilities for request applications within a compact 1 GB memory footprint. This low-latency model, fine-tuned on specific tasks, processes inputs in English, Spanish, Chinese, and Russian while delivering calibrated probabilities without generating text. Key functionalities include tasks like relevance assessment, prompt injection detection, conversation analytics, and intent recognition. This is particularly beneficial for environments lacking GPU support, allowing broader access to advanced AI capabilities. Despite being a smaller version of its predecessor, arc-decide-0.6b-v0.1, the new model meets or exceeds performance benchmarks in several areas, including relevance and grounding, though it slightly lags in BoolQ and deepset injection tasks. Its architecture, utilizing only 16 of the original layers and employing a hybrid approach to training data, achieves comparable accuracy while ensuring efficiency and a lower resource requirement, which is especially appealing for enterprise applications or smaller-scale implementations. This balance of performance and accessibility could drive further adoption of AI decision-making tools in various sectors.