SenseChat-5

SenseNova-V5 (DayDayUp 5.0 / SenseChat V5) is a large-scale hybrid expert language model officially released by SenseTime on April 23, 2024. As a significant milestone in the DayDayUp large-model series, its core positioning is “a Chinese large model at the GPT-4 Turbo level + globally leading humanities capabilities + a balanced and comprehensive intelligence foundation.”

• Breakthrough Comprehensive Chinese Performance: In the authoritative Chinese large-model benchmark SuperCLUE May 2024 Ranking, it achieved a total score of 80.03, setting a new domestic record and surpassing GPT-4-Turbo-0125 (79.8) for the first time, thus topping the list of Chinese-language capabilities.
• World’s No. 1 Humanities Capability: It scored 82.20 points in humanities tasks—including knowledge encyclopedias, long texts, role-playing, semantic understanding, generative creation, and safety alignment—significantly outperforming GPT-4-Turbo (+4.40 points). Several sub-items also set new domestic best records.
• Nationally Leading Science and Technology Capability: In science and technology tasks such as computation (80.6), logical reasoning (73.8), and tool usage (80.8), it scored 76.78 points, ranking first in China and demonstrating well-rounded capabilities.
• Ultra-Large-Scale MoE Architecture: Adopting a hybrid expert (MoE) design with 600 billion parameters, it features efficient activation of parameters and supports an ultra-long context window of 200K tokens, making it ideal for processing lengthy documents and handling complex dialogues.
• High-Quality Data and Training System: Built upon a 10T-token next-generation data pipeline, it integrates hundreds of billions of synthesized chain-of-thought data; employing a multi-stage training pipeline consisting of three-phase pre-training + two-phase SFT + online RLHF, it achieves multi-dimensional alignment with human preferences.

───────────────────────────────────────────────────────────────────

Core Capabilities

📚 Ultra-Long Text Understanding and Generation: Accurately handles documents spanning hundreds of pages, completing tasks such as summarization, question answering, and style continuation (e.g., legal contracts, academic papers).

🎭 High-Fidelity Role-Playing: Supports customized character settings, memory retention, and emotional expression, making it suitable for virtual companionship, educational practice, and digital customer service agents.

🧠 Tool Invocation and Code Generation: Equipped with strong capabilities in function invocation, API usage, and multi-language programming, enabling the construction of automated workflows.

🛡️ Strong Security and Value Alignment: Achieves a score as high as 90.2 in traditional security tasks, effectively preventing the generation of harmful, biased, or non-compliant content.