HyperAI

Latest AI news and updates from around the world

Online Tutorial | DeepSeek-V4 Brings Visual Agent Capabilities, Surges to 36.5 on ApexBench

Currently, the HyperAI tutorial section has launched “Multimodal Reasoning with DeepSeek-V4-Flash-Vision-Exp and Deployment via Open WebUI,” enabling one-click model deployment for rapid exploration of its visual understanding, multimodal agent capabilities, and complex task execution performance.

5 hours ago

DeepSeek

Multimodal

LLM

MiniMax H3 Breaks Through the Boundaries of Video Generation, Generating Audio and Video Content in an Integrated Manner; Ornith-1.5-35B-A3B Explores a New Mode of self-evolutionary Training for Inference models.

8.21-8.27 Weekly Highlights

7 days ago

Command Palette

News

Online Tutorial | DeepSeek-V4 Brings Visual Agent Capabilities, Surges to 36.5 on ApexBench

MiniMax H3 Breaks Through the Boundaries of Video Generation, Generating Audio and Video Content in an Integrated Manner; Ornith-1.5-35B-A3B Explores a New Mode of self-evolutionary Training for Inference models.

27B AI's Ability to Reproduce scientists' Papers Surpasses GPT-5.5/Claude Opus 4.8; Faraday Explores long-term Scientific innovation.

Online Tutorial | Generate a Complete Song in One Sentence: MiniMax Music 3 Handles Composition, Arrangement, and vocals.

Covering 11 Disciplines and 60 Research Tasks, Tsinghua University, MIT, Harvard, and Others Have Proposed the ASI-Bench to Test AI's Independent Research capabilities.

From Code Development to Automated Execution, Muse-Glimmer-30B Makes Local Agents More Powerful; Qwen3.8-27B-FP8 Released, Pushing the Limits of Efficient Inference With 27 Billion parameters.

Analytical Efficiency Improved by Over 200 Times! Argonne National Laboratory Proposes a Nanobeam Diffraction Analysis Method Based on Unsupervised Training, Enabling DONUT to Accurately Extract Material features.

Event Recap | From Compilation Optimization to Scenario Applications: A Multi-Perspective Analysis of AI Compiler Technology Innovation and Implementation Experience in the Agent Era

Online Tutorial | Qwen 3.8-27B: Major Open Source Release, Upgrading Reasoning/Programming/Agent Capabilities

Bi Sheng's Compiler Core, AscendNPU IR, Has Been open-sourced. Huawei Architect Hai Lijuan Explains the Technical Progress and Highlights Its Support for Operator Ecosystems Such As Triton.

128K Long Context + Agent Reinforcement Training! LFM2.5-2.6B Unlocks Efficient Deployment of Large Models on Edge; DETR Uses Transformer to Cut Through NMS and Anchors, Reshaping Object Detection

Online Tutorial | 35k Stars Achieved: HKU Team Open-Sources DeepTutor, a Single Agent Connecting Knowledge Q&A, Interactive Learning, Practice Generation, and Deep Research

Leveraging the low-complexity Matrix Multiplication Principles of Strassen and LCMA, Tencent FalconGEMM Explores Matrix Multiplication Optimizations That Surpass Hardware limitations.

Speed Equals Intelligence! TileRT Team Member Ma Lingxiao Explains: Computational Exploration and Collaborative Design for ultra-low Latency Large Model inference.

Chen Tianqi: In the Agent Era, How Should We Rethink AI Compilers?

Online Tutorial | ProteinGym #1! VenusREM Uses "search Enhancement" to Predict the Impact of Protein Mutations and Accelerate Protein design.

The Beijing Academy of Artificial Intelligence (BAAI) Proposed the Triton-TLE Layered Language Extension and Achieved a Hundredfold Automatic Tuning Speedup Through FlagTree Compilation optimization.

The First Standardized Benchmark for Protein Mutation Ranking in LLM! A Harvard University Team Proposes PG-LLM, Which Evaluates 13 Mainstream Models and 95 Specialized Models in One go.

Online Tutorial | Achieving Raw Image Editing With 4B Parameters: Microsoft Open-Sources Mage-Flow, Enabling Second-Level Image Inference and Supporting Up to 2048 Resolution

The U.S. Geological Survey Has Proposed the DAS Earthquake Early Warning Model, Which Uses 4-second P-wave Strain Data to Rapidly Classify Moderate to Strong earthquakes.

One Day to Go! Zhiyuan, TileRT, Tencent, Huawei, and Zhiyuan Innovation Gather to Explore multi-level Collaborative Optimization of AI compilation.

MiniCPM-RobotManip, With Its 1.5 Billion Parameters, Is Officially Open-sourced; NVIDIA Releases the full-modal Model Cosmos3-Edge, Covering Text, Images, Videos, and Motion sequences.

Dataset Compilation | From Competition Math to Tool Applications: 9 Open-Source Math Datasets From MIT, NVIDIA, Huazhong University of Science and Technology, etc., Covering CoT, Multimodal Reasoning, and Long-Chain Thinking Training

Online Tutorial | Run Agents Without Billions of Parameters! Boss Zhipin's Nanbeige Lab Open Sources Nanbeige 4.2-3B, Giving Small Models a "Brain"

Based on Large Model Inference and the Use of the MCP Tool, AI X-ray Scientists at Stanford University Autonomously Completed single-crystal Diffraction Alignment at a Synchrotron Radiation source.

Text/LaTeX/HTML Tables All in One Step! OvisOCR2 Achieves end-to-end Intelligent Document Parsing; 14,000+ Element Annotations, Tens of Thousands of Language Commands! Voxel51 Releases the SceneFun3D Indoor Scene micro-interaction dataset.

Online Tutorial | Compress a 27B Large Model to 7.2GB! Ternary-Bonsai Uses "ternary Magic" to Make Large Models Run on Personal computers.

Argonne National Laboratory in the United States Proposed ChemGraph, Which Uses 13 Benchmark Tests to Evaluate the Value of Agents in the Field of Computational chemistry.

Online Tutorial | Load ComfyUI Workflow With One Click: Master AI Drawing Without Writing a Single Line of Code

Based on Training With Over 10,000 Tumor Samples, Harvard Medical School and Others Proposed COMPASS, a pan-cancer Basic Model, Which Outperforms 22 Existing Methods on average.

Online Tutorial | DeepSeek-V4 Brings Visual Agent Capabilities, Surges to 36.5 on ApexBench

MiniMax H3 Breaks Through the Boundaries of Video Generation, Generating Audio and Video Content in an Integrated Manner; Ornith-1.5-35B-A3B Explores a New Mode of self-evolutionary Training for Inference models.

27B AI's Ability to Reproduce scientists' Papers Surpasses GPT-5.5/Claude Opus 4.8; Faraday Explores long-term Scientific innovation.

Online Tutorial | Generate a Complete Song in One Sentence: MiniMax Music 3 Handles Composition, Arrangement, and vocals.

Covering 11 Disciplines and 60 Research Tasks, Tsinghua University, MIT, Harvard, and Others Have Proposed the ASI-Bench to Test AI's Independent Research capabilities.

From Code Development to Automated Execution, Muse-Glimmer-30B Makes Local Agents More Powerful; Qwen3.8-27B-FP8 Released, Pushing the Limits of Efficient Inference With 27 Billion parameters.

Analytical Efficiency Improved by Over 200 Times! Argonne National Laboratory Proposes a Nanobeam Diffraction Analysis Method Based on Unsupervised Training, Enabling DONUT to Accurately Extract Material features.

Event Recap | From Compilation Optimization to Scenario Applications: A Multi-Perspective Analysis of AI Compiler Technology Innovation and Implementation Experience in the Agent Era

Online Tutorial | Qwen 3.8-27B: Major Open Source Release, Upgrading Reasoning/Programming/Agent Capabilities

Bi Sheng's Compiler Core, AscendNPU IR, Has Been open-sourced. Huawei Architect Hai Lijuan Explains the Technical Progress and Highlights Its Support for Operator Ecosystems Such As Triton.

128K Long Context + Agent Reinforcement Training! LFM2.5-2.6B Unlocks Efficient Deployment of Large Models on Edge; DETR Uses Transformer to Cut Through NMS and Anchors, Reshaping Object Detection

Online Tutorial | 35k Stars Achieved: HKU Team Open-Sources DeepTutor, a Single Agent Connecting Knowledge Q&A, Interactive Learning, Practice Generation, and Deep Research

Leveraging the low-complexity Matrix Multiplication Principles of Strassen and LCMA, Tencent FalconGEMM Explores Matrix Multiplication Optimizations That Surpass Hardware limitations.

Speed Equals Intelligence! TileRT Team Member Ma Lingxiao Explains: Computational Exploration and Collaborative Design for ultra-low Latency Large Model inference.

Chen Tianqi: In the Agent Era, How Should We Rethink AI Compilers?

Online Tutorial | ProteinGym #1! VenusREM Uses "search Enhancement" to Predict the Impact of Protein Mutations and Accelerate Protein design.

The Beijing Academy of Artificial Intelligence (BAAI) Proposed the Triton-TLE Layered Language Extension and Achieved a Hundredfold Automatic Tuning Speedup Through FlagTree Compilation optimization.

The First Standardized Benchmark for Protein Mutation Ranking in LLM! A Harvard University Team Proposes PG-LLM, Which Evaluates 13 Mainstream Models and 95 Specialized Models in One go.

Online Tutorial | Achieving Raw Image Editing With 4B Parameters: Microsoft Open-Sources Mage-Flow, Enabling Second-Level Image Inference and Supporting Up to 2048 Resolution

The U.S. Geological Survey Has Proposed the DAS Earthquake Early Warning Model, Which Uses 4-second P-wave Strain Data to Rapidly Classify Moderate to Strong earthquakes.

One Day to Go! Zhiyuan, TileRT, Tencent, Huawei, and Zhiyuan Innovation Gather to Explore multi-level Collaborative Optimization of AI compilation.

MiniCPM-RobotManip, With Its 1.5 Billion Parameters, Is Officially Open-sourced; NVIDIA Releases the full-modal Model Cosmos3-Edge, Covering Text, Images, Videos, and Motion sequences.

Dataset Compilation | From Competition Math to Tool Applications: 9 Open-Source Math Datasets From MIT, NVIDIA, Huazhong University of Science and Technology, etc., Covering CoT, Multimodal Reasoning, and Long-Chain Thinking Training

Online Tutorial | Run Agents Without Billions of Parameters! Boss Zhipin's Nanbeige Lab Open Sources Nanbeige 4.2-3B, Giving Small Models a "Brain"

Based on Large Model Inference and the Use of the MCP Tool, AI X-ray Scientists at Stanford University Autonomously Completed single-crystal Diffraction Alignment at a Synchrotron Radiation source.

Text/LaTeX/HTML Tables All in One Step! OvisOCR2 Achieves end-to-end Intelligent Document Parsing; 14,000+ Element Annotations, Tens of Thousands of Language Commands! Voxel51 Releases the SceneFun3D Indoor Scene micro-interaction dataset.

Online Tutorial | Compress a 27B Large Model to 7.2GB! Ternary-Bonsai Uses "ternary Magic" to Make Large Models Run on Personal computers.

Argonne National Laboratory in the United States Proposed ChemGraph, Which Uses 13 Benchmark Tests to Evaluate the Value of Agents in the Field of Computational chemistry.

Online Tutorial | Load ComfyUI Workflow With One Click: Master AI Drawing Without Writing a Single Line of Code

Based on Training With Over 10,000 Tumor Samples, Harvard Medical School and Others Proposed COMPASS, a pan-cancer Basic Model, Which Outperforms 22 Existing Methods on average.