SpaceXAI 推出 Grok 4.6 模型
Related
Github Copilot integrates the open-source model Kimi K2.7.
According to Mars Finance, Dark Side of the Moon announced that Github Copilot has integrated its first open-source model, Kimi K2.7.
Grok Build integrates with Composer 2.5: xAI terminal agents introduce the Cursor main model.
According to Beating, xAI, Elon Musk's AI startup, announced that its command-line agent development tool Grok Build has fully integrated Composer 2.5, a top-tier programming model developed by Cursor. Developers subscribing to X Premium+ or SuperGrok can directly switch to Composer 2.5 in the terminal menu. Cursor released Composer 2.5, a high-performance programming model, on May 18th, focusing on high-precision multi-file editing and long-term instruction compliance. The model's underlying architecture uses the commercially licensed open-source Kimi K2.5 from Dark Side of the Moon, and was trained using large-scale reinforcement learning and synthetic task distillation by the Cursor team. The "local reinforcement learning with text feedback" technique introduced during training provides targeted hints at precise points where tool calls fail. The model can automatically correct local errors, improving the success rate of cross-file code refactoring and long-term execution. Official demonstrations show that when faced with natural language commands such as "improve the mobile responsiveness of the game loading menu," the agent can autonomously retrieve relevant code. After performing regular expression searches on files such as useMobile.ts and Game.tsx, the agent directly locates and modifies the specific implementation of localStorage in page.tsx. The interface supports shortcut commands such as `/fork` (clone the current session to a parallel sub-agent), `/compact` (compress history to reduce token consumption), and `/copy` (copy the latest answer), and provides an "always-approve" automatic authorization mode. The terminal ecosystem integration stems from the computing power sharing cooperation between Cursor and SpaceXAI. Currently, Cursor has access to computing resources, including the Colossus 2 super cluster, for model training. The introduction of Composer 2.5 is the first implementation of the cooperation between the two parties in the xAI terminal tool.
Microsoft will release its self-developed programming model next week to counter Cursor and Claude's encroachment on Copilot.
According to Beating, Microsoft plans to release several self-developed AI models at its annual Build developer conference in San Francisco next week. To reclaim the AI coding market share eroded by competitors, Microsoft is focusing on launching its self-developed programming models to enhance the competitiveness of GitHub Copilot and counter the market encroachment brought by Cursor and Claude Code. In addition to programming models, Microsoft is also planning to launch self-developed large models with various parameter specifications, covering vertical tasks such as transcription, reasoning, speech, and image processing. The research team under Microsoft AI CEO Mustafa Suleyman has not had a model reach the top of mainstream industry leaderboards in the past two years; this release will be the first major test of the team's self-developed capabilities. Previously, due to an exclusive agreement with OpenAI restricting Microsoft's rights to develop human-level AI until April 2026, its self-development work was once limited; it wasn't until April of this year, when the two companies revised their cooperation terms, that Microsoft regained its research freedom. In its software services, Microsoft has long relied on OpenAI's technology, which is licensed free of charge until 2032, and the Anthropic model, which it purchased for a fee, to power GitHub Copilot and 365 Copilot. The high cost of external procurement has forced Microsoft to raise the price of some GitHub Copilot packages and impose quotas on the use of the Anthropic model. By distributing lower-cost, self-developed models through the Azure cloud platform, Microsoft can provide developers with a cheaper alternative to Claude, thereby reducing the daily operating costs of AI features in Office software. More importantly, when the free license for OpenAI expires in about six years, self-developed large models will become a key asset for Microsoft to break free from expensive suppliers and maintain its core independence.
Grok's gray-scale testing user memory feature supports viewing and editing user profiles.
According to Beating's monitoring, xAI's large model Grok is currently undergoing gray-scale testing of a new user memory feature. The test page shows that the memory feature automatically records key personal information mentioned by users in conversations and generates a systematic user profile. Users can directly view and edit Grok's recorded personal memory files, providing continuous context for subsequent conversations and improving the smoothness of interaction. Leaked screenshots show that Grok's automatically generated user profiles include multiple dimensions such as social media accounts, hardware configurations, daily occupations, and subscription versions. It can even automatically identify a user's Discord community administrator status and GitHub open-source project details based on chat history. In terms of product form, Grok's user profile mechanism is similar to OpenAI's previously launched ChatGPT memory feature, but Grok's organization dimensions are more specific and structured, using a list format to categorize users' personal backgrounds and technology stacks in detail. Currently, xAI has not yet announced the official launch date or specific rollout scope of the memory feature.
Grafana Labs' GitHub repository was attacked by ransomware; customer systems were unaffected.
According to ChainCatcher, Grafana Labs' official blog disclosed that on May 16, 2026, Grafana Labs confirmed it was targeted by a cybercrime group. Attackers exploited a TanStack npm supply chain vulnerability (Mini Shai-Hulud campaign) to compromise its GitHub repository on May 11, downloading public and private source code, internal operational information, and emails from some business contacts, before issuing ransom threats. Grafana Labs stated that the codebase was only downloaded, not modified, and that customer production systems and the Grafana Cloud platform were unaffected. The company has decided not to pay the ransom and has reported the incident to federal law enforcement. The company has initiated measures including token rotation, log review, audit submission, and CI/CD pipeline security hardening. A full post-incident report will be released after the investigation is completed.
Musk: SpaceX AI will release Grok 4.5 tomorrow.
According to Foresight News , Musk announced that SpaceX AI will release Grok 4.5 to the public tomorrow. It is an Opus-level model, but faster, more efficient, and cheaper.