Key takeaways

• DeepSeek V4 Pro is a large-scale Mixture-of-Experts (MoE) model with 1.6T total parameters and 49B activated parameters, designed for advanced reasoning, coding, and complex problem-solving tasks, supporting a 1M-token context window.
• The model achieved a perfect 100% accuracy on the 'Truth-Tellers and Liars Logic Puzzles' benchmark, demonstrating exceptional logical consistency, robust logical deduction, and the ability to handle nested beliefs and complex statements without error.
• DeepSeek V4 Pro demonstrated a significant weakness in an operating system environment, achieving a very low success rate of 2.86% on the AgentBench OS benchmark. It struggled with complex multi-step problem-solving, dynamic OS environment interactions, and precise application of Linux utilities.
• Key limitations include difficulties with complex command chaining, nuanced instruction interpretation, calculations, dynamic filtering, and advanced text processing using bash, often failing to construct correct command sequences for tasks like log analysis.
• While proficient in formal logic and precise deduction, its poor performance on AgentBench OS highlights a gap in its ability to translate high-level reasoning into effective and accurate command sequences in an interactive computing environment.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~1600B

Open weights

Yes

Licence

MIT

Regions

China

Vendor

DeepSeek

Provider model key

deepseek/deepseek-v4-pro