DeepSeek previews new AI model that ‘closes the gap’ with frontier models
DeepSeek previews new AI model that ‘closes the gap’ with frontier models
DeepSeek previewed two new models and said architectural changes make them more efficient and higher-performing than DeepSeek V3.2, while nearly closing the gap with leading models on reasoning benchmarks. The RSS snippet discloses only that there are two models and that they outperform V3.2; model names, parameter counts, benchmark scores, test sets, and release timing are not disclosed. The key question is reproducible evals, because “closes the gap” comes without numbers.
Why it matters: DeepSeek posted two V4 preview models on Hugging Face, claiming a leaner architecture than V3.2 and higher scores, with parameter counts and benchmark numbers not disclosed.