Moonshot AI (Kimi) released its flagship K3 model on July 16. With 2.8 trillion total parameters, it uses a Stable LatentMoE architecture with 896 experts, activating 16 at a time, and natively supports visual understanding and a 1 million token context window. K3 introduces KDA (Kimi Delta Attention) hybrid linear attention, Attention Residuals, and a Per-Head Muon optimizer. The company claims overall scaling efficiency is improved by about 2.5x compared to K2. While K3’s overall performance lags behind Claude Fable 5 and GPT-5.6 Sol, it has surpassed these closed-source models in programming and long-horizon agent tasks. K3 topped the Frontend Code Arena with a score of 1679, surpassing Fable 5 (jumping from 18th place with K2.6 to 1st). It also ranked first in Program Bench (77.8), SWE Marathon (42.0), BrowseComp (91.2), and Automation Bench. K3 demonstrated long-horizon autonomous capabilities: it ran continuously for 24 hours in a sandbox writing GPU kernels, developed a MiniTriton compiler from scratch similar to Triton, and spent 48 hours designing and verifying a 4mm\u00b2 chip. Full model weights will be open-sourced before July 27. API pricing is set at \u00a52 per million tokens for cached input, \u00a520 for uncached input, and \u00a5100 for output.<\/p>\n
Sources:<\/p>\n
On the eve of the official opening of WAIC in Shanghai, 29 countries signed an agreement to establish the World Artificial Intelligence Cooperation Organization (WAICO). If it operates successfully, this will be the world’s first intergovernmental international organization dedicated to AI. The UN Secretary-General, the President of Kazakhstan, the Prime Minister of Thailand, the Prime Minister of Cambodia, and other foreign dignitaries have arrived in Shanghai. Many bloggers view this event as a landmark shift in AI competition from the model layer to the rule-making layer.<\/p>\n
Sources:<\/p>\n
A hacker breached Suno using a worm named Shai-Hulud, obtaining source code related to training data. The leaked information shows Suno scraped approximately 114,000 hours of content from YouTube Music, about 62,000 hours from Pond5, around 12,000 hours from Deezer, and planned to download about 1 million hours of podcasts. The hacker also obtained Suno user emails, phone numbers, and Stripe-related information, affecting hundreds of thousands of users. Suno stated the security incident was discovered and contained in November 2025, with limited impact, and thus users were not notified individually.<\/p>\n
Sources:<\/p>\n
Alaya World has open-sourced its inference code and 15B model weights, supporting a persistent, real-time generative game world. Its core feature is using language to directly manipulate the environment\u2014issuing a command in an icy wasteland causes the ground to crack and skeletons to emerge; saying “make it night” transitions the next scene to darkness. The project uses 3D Cache plus compressed frame history technology to achieve continuous generation exceeding 1 minute while maintaining spatial consistency. Several bloggers who tested it believe future game interaction may shift from fixed skill buttons to natural language-driven world manipulation.<\/p>\n
Sources:<\/p>\n
The launch of Starship Flight 13 was automatically aborted after some engines failed to start, and propellant offloading has begun. SpaceX will remove and replace two Raptor engines, with the next launch attempt expected as early as next week. Musk shared the second episode of SpaceX’s Starship V3 documentary. On the AI front, Grok Build received a major update (v0.2.102) adding conversation navigation and enterprise control features; SuperGrok Heavy now includes X Premium+. For Starlink, V3 satellites are expected to increase space-based bandwidth by about two orders of magnitude; Starlink also launched in C\u00f4te d’Ivoire, and Tesla officially entered Uruguay. Musk also shared community feedback stating Grok 4.5 is the first Grok model capable of daily development use.<\/p>\n
Sources:<\/p>\n
From OpenAI, the ChatGPT desktop app received a significant update: chat history and projects now appear in the sidebar, switching between Chat and Work modes is clearer, and the user experience is converging across desktop, web, and mobile. Codex launched a PR Chat feature, allowing pull request review and inline editing within Codex. Codex Micro sold out within 14 hours of release. OpenAI CEO Sam Altman acknowledged the past 12 months haven’t been the company’s best but stated the team is doing amazing work, the next 12 months will be the best yet, and increasingly users are interacting with ChatGPT via voice rather than text. Several bloggers noted Codex’s new layout resembles Claude’s desktop style and that a new referral program has started.<\/p>\n
Sources:<\/p>\n
NVIDIA introduced RoboTTT, extending robot visuomotor context from single-step or a few frames to 8000 timesteps (about 5 minutes at 30Hz). It successfully completed a full 10-stage assembly task on a physical robot and can automatically correct errors. This direction is considered a potential third scaling curve for robot foundation models, following parameter scale and data volume.<\/p>\n
Additionally, LiveKit optimized for Gemma 4.31B: in real-time speech scenarios, first-token latency is only 192ms (compared to GPT-4.1’s 1006ms) at about one-sixth the cost, with an 88% task completion rate in hotel front desk tests. Browser Use added a video recording feature where an agent can record while operating, with AI automatically editing it into a tutorial video.<\/p>\n
Sources:<\/p>\n
Statistics:<\/p>\n
Kimi releases the world’s first open-source 3 trillion parameter model, surpassing closed-source rivals in programming. WAIC establishes the first intergovernmental AI organization, signaling a shift to rule-making competition. Suno’s\u2026<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1469","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1469","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1469"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1469\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1469"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1469"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1469"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}