Jiwo 0.8B – Pushing the sub 1B decision models leaderboard (huggingface.co)

The Jiwo 0.8B model has recently achieved groundbreaking results, earning the highest Decision Index score (28.00) among decision models with fewer than 1 billion parameters, ranking first out of 33 on the live leaderboard. This model differentiates itself by processing a state alongside multiple typed questions in a single forward pass, specifically focused on return probabilities rather than text generation. Its architecture includes a base model from Qwen 3.5-0.8B with full weighted components, supporting various question types such as choices, scores, and true/false inquiries, and it operates with a training context length of 8,192 tokens. This achievement signals a significant advancement for the AI/ML community, particularly in the realm of decision-making processes. The model demonstrated a notable performance improvement of over 21 points compared to an untrained version of the base model and outperformed the best-known fine-tuned alternatives. With a median request latency of just 48 ms, Jiwo 0.8B illustrates both speed and efficiency, making it a formidable addition to decision-making AI tools. Importantly, developers can easily implement Jiwo 0.8B using a straightforward API, opening the door for a variety of practical applications in sectors where swift and accurate decision-making is critical.