TLDR: DeepSeek has launched DeepSeek-V3.1-Terminus, an updated AI model focused on enhancing AI agent functionalities. This iteration brings significant improvements in coding, search capabilities, and language consistency, addressing prior user feedback. Built on a hybrid Mixture of Experts (MoE) architecture with 685 billion parameters, the model shows strong performance across various benchmarks and is available to developers on open-source platforms.
Hangzhou, China – DeepSeek, a prominent artificial intelligence startup, announced today the release of its updated foundation model, DeepSeek-V3.1-Terminus. Launched on September 23, 2025, this new iteration marks a strategic advancement in the company’s focus on AI agents, aiming to empower software that can automate complex tasks for users.
DeepSeek-V3.1-Terminus is an iterative enhancement to its predecessor, DeepSeek-V3.1, which was previously recognized by Artificial Analysis as the company’s most advanced model. The latest update specifically targets user-reported issues and amplifies core strengths, particularly in agentic capabilities.
Key improvements in DeepSeek-V3.1-Terminus include enhanced coding and search functionalities. The model also boasts significantly improved language consistency, directly addressing earlier user feedback that highlighted instances of the DeepSeek chatbot producing illegible symbols and unprompted switches between Chinese and English in its responses. This refinement ensures more stable and reliable outputs for real-world applications.
Architecturally, DeepSeek-V3.1-Terminus retains the hybrid Mixture of Experts (MoE) framework, a design choice that combines dense and sparse components. This allows the model to activate only the most relevant expert modules for specific tasks, leading to high efficiency and reduced computational overhead compared to fully dense models. The model is built upon a substantial 685 billion parameters distributed across these expert modules.
In terms of performance, DeepSeek-V3.1-Terminus has shown slight improvements on several popular benchmarks. These include “Humanity’s Last Exam,” a rigorous academic test designed to push the limits of AI systems, as well as various coding benchmarks. Experts widely consider strong coding abilities crucial for developing robust AI systems with general capabilities. The model also demonstrated improvements on the OpenAI-backed BrowseComp benchmark, which evaluates an AI’s ability to retrieve challenging information from the internet. However, it experienced a minor decrease in its score on the Chinese-language version of this test, falling from 49.2% to 45%.
Also Read:
- Replit Unveils Agent 3: A New Era of Autonomous AI-Powered Software Development
- New Open-Source Cybersecurity AI Framework Launched to Empower Security Teams
DeepSeek has made the weights for DeepSeek-V3.1-Terminus available on prominent open-source platforms such as Hugging Face and Alibaba-backed ModelScope. This move enables global developers to download and build upon the model, fostering broader innovation within the AI community. The company continues to attract substantial global interest, with Clément Delangue, CEO of Hugging Face, noting DeepSeek’s poised position to surpass 100,000 followers on the platform, underscoring its growing influence in the AI landscape.
