AI21: Jamba Large 1.7

AI21: Jamba Large 1.7, created on August 8, 2025, is a hybrid SSM-Transformer model with a substantial 256K context window, designed for improved grounding, instruction-following, and efficiency. In terms of overall performance, it performs among the fastest models, ranking in the 64th percentile for speed across eight benchmarks. Its pricing is moderate, placing it in the 24th percentile. Notably, the model demonstrates strong reliability with a 92% success rate, indicating consistent operational stability. Analyzing benchmark results, Jamba Large 1.7 exhibits exceptional performance in Ethics and Email Classification, achieving 99% accuracy in both, placing it in the 51st and 79th percentiles respectively. General Knowledge is also a strength at 93% accuracy. However, a significant weakness is its high hallucination rate, with only 52% accuracy in acknowledging uncertainty, placing it in the 7th percentile. Performance in Mathematics (40% accuracy) and Instruction Following (44% accuracy) is also below average. Reasoning and Coding show moderate performance at 54% and 67% accuracy, respectively. The model's key strengths lie in its ethical adherence, classification capabilities, and overall reliability, while its primary areas for improvement are reducing hallucinations and enhancing mathematical and instruction-following precision.