AI 新闻 —— July 16, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Moonshot AI launches Kimi K3, a 2.8T parameter LLM with higher pricing
Moonshot AI has launched Kimi K3, a new large language model boasting 2.8 trillion parameters and described as their most capable model to date. The model is available via API and website, with an open-weight release anticipated by July 27, 2026. Kimi K3 shows strong performance…
-
Dutch Navy integrates AI and drones to transform sea defenses · 2 sources tracked
The Royal Netherlands Navy is integrating AI and uncrewed systems to enhance its sea defenses, aiming for over half of its operations to be conducted by these technologies within five years. This initiative, led by Captain Sjoerd Feenstra, focuses on using autonomous drones, und…
-
Meta releases Muse Spark 1.1 with 1M context window and competitive pricing
Meta has released Muse Spark 1.1 into public preview, a multimodal model that offers a 1 million token context window and a cost-effective pricing structure. The model is priced at $1.25 per million input tokens and $4.25 per million output tokens, with an output speed of 118.1 …
-
Moonshot AI's Kimi 3 aims to rival Anthropic's Opus 4.8, eyes $31.5B valuation · 2 sources tracked
Moonshot AI is reportedly developing its Kimi 3 model, aiming to match or exceed the performance of Anthropic's Opus 4.8. This new model is expected to be the largest open-weight AI model from China, featuring a parameter count between 2 and 3 trillion. The company is also seeki…
-
Moonshot AI launches 2.8T parameter Kimi K3 model with 1M context window
Moonshot AI has launched Kimi K3, a 2.8 trillion parameter model featuring Kimi Delta Attention and Attention Residuals. This model natively supports visual understanding and boasts a 1 million token context window, making it the first open-source model of its scale. Kimi K3 is …
-
Applied Computing raises $20M for AI model targeting oil and gas plants
Applied Computing, a London-based startup, has secured $20 million in Series A funding to develop its foundation AI model, Orbital, for the oil, gas, and petrochemical industries. Orbital integrates time-series, physics-based, and language models to analyze sensor data, predict …
-
xAI launches Grok 4.5 with new automation features and pricing
xAI has launched new automation features for its Grok AI, making them available on the Grok website and mobile applications. The Grok 4.5 model is now accessible through Grok Build, Cursor, and the xAI console, with pricing set at $2 per million input tokens and $6 per million o…
-
Moonshot AI previews Kimi K3, challenging Fable 5 in video generation
Moonshot AI's new model, Kimi K3, has been previewed, with benchmark videos circulating on social media. Early tests suggest K3's capabilities in video generation detail and animation smoothness are comparable to Fable 5. The model's appearance on the Arena benchmark indicates i…
-
Thinking Machines releases Inkling, a 975B parameter open-weights multimodal model
Thinking Machines Lab has released Inkling, a new family of open-weights multimodal foundation models. The flagship Inkling model features 975 billion total parameters with 41 billion active, supports a 1 million token context window, and is trained on 45 trillion tokens of text…
-
Moonshot AI releases Kimi K3, world's first open-source 3T parameter model
Moonshot AI has officially launched Kimi K3, an open-source model with 2.8 trillion parameters. This model utilizes Kimi Delta Attention and Attention Residuals technologies, natively supports visual understanding, and boasts a 1 million token context window. Designed for advanc…
-
xAI prices Grok 4.5 at $2/$6 per million tokens
xAI has announced pricing for its Grok 4.5 model, which will cost $2 per million input tokens and $6 per million output tokens. The model is now available on the Grok website and through its applications on iOS and Android. Additionally, Grok 4.5 is accessible via Grok Build, Cu…
-
AI tutor ensures circuit simulation accuracy via rigorous verification
A new AI tutor for circuit simulation has been developed, focusing on ensuring accuracy through a robust verification system. Instead of directly generating circuit layouts, the AI proposes logical netlists, which are then processed by a separate engine. This engine verifies the…
-
RAG systems introduce new attack surfaces beyond prompt injection
Retrieval-Augmented Generation (RAG) systems introduce significant security vulnerabilities beyond traditional prompt injection. Attackers can exploit the document ingestion pipeline, vector storage, and retrieval assembly processes to manipulate LLM outputs. Treating all retrie…
-
Classical ML methods show promise in detecting LLM-generated text
Researchers are exploring the use of traditional machine learning models to detect text generated by large language models (LLMs). These classical methods, such as Support Vector Machines and Naive Bayes classifiers, offer advantages in interpretability and efficiency compared t…
-
Former OpenAI CTO releases open-weights LLM, criticizes closed models
Thinking Machines, a company founded by former OpenAI CTO Greg Brockman, has released an open-weights model named Thinking Machines. This model boasts 975 billion parameters and is positioned as an alternative to Chinese large language models. The release aims to provide a more …
-
Google stealthily updates Gemma 4 model, fixing bugs and improving performance
Google has released an update for its open AI model, Gemma 4. This update reportedly improves performance on Nvidia Hopper GPUs and resolves issues with tool calling bugs and truncated responses. The changes were implemented without a formal announcement, appearing as a stealth …
-
Prompt chaining improves LLM accuracy by breaking tasks into smaller steps
Prompt chaining, a technique inspired by Unix pipelines, offers a more reliable method for complex tasks than using a single, large prompt with Large Language Models. This approach breaks down a task into a sequence of smaller, single-purpose prompts, where the output of one pro…
-
Anthropic's Claude introduces built-in memory, with external options for deeper recall
Anthropic's Claude now features a built-in memory that retains user preferences and light context across conversations, a significant improvement over its previous stateless interactions. However, this built-in memory has limitations, primarily storing only superficial informati…
-
AI Model Servers Show Systemic Security Flaws, Critical Vulnerabilities Found
A security audit of over 50 open-source Model Context Protocol (MCP) servers revealed systemic vulnerabilities, with over 60% exhibiting unsafe patterns. Researchers identified critical flaws, including remote code execution with cloud credential theft and chained vulnerabilitie…
-
AI tools marketed to police raise transparency and accountability concerns
The AI industry is increasingly targeting law enforcement with a range of automated tools, promising efficiency and improved decision-making. Products like facial recognition, automated license plate readers, and predictive policing software are being marketed to police departme…