Key takeaways

• DeepSeek V3.1 is a 671B parameter hybrid reasoning model (37B active) designed for complex text generation and reasoning, supporting both 'thinking' and 'non-thinking' modes via prompt templates. It aims to solve problems requiring strong mathematical reasoning and precise calculations.
• It is a Transformer-based model with open weights, extending DeepSeek-V3 base. It was trained with a two-phase long-context training process, supporting up to 128K tokens, and uses FP8 microscaling for efficient inference.
• The model excels in mathematical reasoning, precise calculations, and robust factual recall, with strong adherence to formatting (e.g., LaTeX). Known weaknesses include occasional errors in complex logical reasoning, minor arithmetic errors, and difficulties with specific geometry interpretations or ambiguous problem constraints.
• DeepSeek V3.1 achieved 93.8% accuracy on the 'MATH-500' dataset, demonstrating strong problem-solving capabilities across diverse mathematical domains while providing detailed, step-by-step explanations.
• The model struggled with the 'Humanity's Last Exam' dataset (biology/medicine), achieving only 5.86% accuracy, indicating difficulties with nuanced interpretation, integrating multiple complex constraints, and applying real-world contextual factors.

Technical specifications

Context length

164k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~685B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

DeepSeek

Provider model key

deepseek/deepseek-chat-v3.1