DeepSeek-V3.1-Terminus Launches with Significant Performance Improvements and Enhanced Deep Reasoning Capabilities

[AIbase Report] DeepSeek has released the latest upgraded DeepSeek-V3.1-Terminus model on its official API platform tonight, and subsequently announced that the model will be open-sourced. According to the official documentation, the new model maintains its original powerful capabilities, fixes issues such as language inconsistency and occasional abnormal characters that appeared after the release of DeepSeek-V3.1, and further optimizes the performance of programming and search agents.

According to benchmark test data published by the official WeChat account of DeepSeek, in non-Agent (agent) tests, the performance of DeepSeek-V3.1-Terminus has improved by 0.2% to 36.5% compared to the previous version. Among these, the performance improvement in HLE (Human Ultimate Test) is the most significant, a test that mainly evaluates the model's ability in expert-level high-difficulty knowledge, multimodal content, and deep reasoning.

In addition, in Agent evaluations, DeepSeek-V3.1-Terminus also shows slight improvements in web browsing, simple Q&A, and multiple programming tests.

DeepSeek

Currently, the official DeepSeek App, web version, mini program, and DeepSeek API have all been updated to the DeepSeek-V3.1-Terminus model.

Open Source Address:

https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus

https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Terminus

Ant Group Opensources the World's First Trillion-Parameter Large Model Ring-1T-preview with Code Generation Capabilities Exceeding GPT-5

Ant Group opensources the trillion-parameter inference large model Ring-1T-preview, the world's first open-source trillion-parameter inference model. The preview version shows outstanding performance in natural language reasoning, achieving a score of 92.6 on AIME25, surpassing all known open-source models such as Gemini 2.5 Pro, and approaching GPT-5's score of 94.6; it also performed well on CodeForces tests.

Alibaba Cloud Launches the World's First All-Modal AI Model Qwen3-Omni, Enabling Unified Processing of Text, Images, Audio, and Video

Alibaba Cloud released the world's first native end-to-end all-modal AI model Qwen3-Omni and open-sourced it. The model supports multi-modal inputs such as text, images, audio, and video, and enables real-time streaming output with fast response. Through text pre-training and multi-modal mixed training, Qwen3-Omni possesses strong cross-modal capabilities and demonstrates advanced performance in multiple fields.

Tongyi DeepResearch Launches! Fully Open-Source AI Model Makes Research Simpler

In the field of artificial intelligence, the latest research results released by the Tongyi DeepResearch team have attracted widespread attention. This breakthrough not only elevates AI from 'being able to chat' to 'being able to conduct research', but also demonstrates its outstanding performance in an open manner. Tongyi DeepResearch has achieved state-of-the-art results in multiple authoritative benchmark tests, with overall capabilities even surpassing many internationally renowned models. Moreover, the model, framework, and solutions are fully open-sourced, truly bringing the productivity of deep research to the world.

Groq Raises $750 Million in Funding, Valuation Reaches $6.9 Billion

Recently, Groq, a pioneer in AI inference technology, announced the successful fundraising of $750 million, with its valuation reaching $6.9 billion after the funding. This round was led by the Texas-based investment company Disruptive, and it also received significant investments from Blackstone, Newport Group, Deutsche Telekom Capital Partners, and a major U.S. West Coast fund manager. Additionally, companies such as Samsung, Cisco, D1, Altimeter, 1789Capital, and Infinitum continued to support the company.

Redefining Tradition! Mini-o3 Open-Source Model Achieves Ultra-Long Visual Reasoning, Deep Thinking Is No Longer a Challenge

Recently, ByteDance and the University of Hong Kong jointly launched a new open-source visual reasoning model - Mini-o3, marking another major breakthrough in multi-turn visual reasoning technology. Unlike previous visual language models (VLMs) that could only conduct 1-2 rounds of dialogue, Mini-o3 limited the number of dialogue rounds to 6 during training, but during testing it can extend the reasoning rounds to dozens, greatly enhancing the ability to handle visual questions. The strength of Mini-o3 lies in its deep reasoning in high-difficulty visual search tasks, reaching