Deep Cogito: Cogito V2 Preview Deepseek 671B

Deep Cogito's Cogito V2 Preview Deepseek 671B, a multilingual MoE model, demonstrates exceptional speed, consistently ranking among the fastest models across seven benchmarks. Its pricing is moderate, positioned at the 36th percentile. The model's core strengths lie in its advanced reasoning and mathematical capabilities. It achieved an impressive 88.0% accuracy in Reasoning, placing it in the 78th percentile, and an outstanding 94.0% accuracy in Mathematics, ranking in the 89th percentile. These results highlight its proficiency in complex problem-solving and quantitative tasks, aligning with its optimization for STEM fields. However, the model exhibits significant weaknesses in other critical areas. It scored 0.0% accuracy across Instruction Following, Coding, General Knowledge, Email Classification, and Ethics benchmarks. This indicates a substantial gap in its ability to follow instructions, perform coding tasks, recall general information, classify emails accurately, and navigate ethical scenarios. While optimized for STEM and programming, the benchmark results for coding do not reflect this optimization. The model's hybrid reasoning via Iterated Distillation and Amplification (IDA) appears to be highly effective for specific analytical tasks but does not translate to broader instruction adherence or factual recall in its current preview state.