Kimi K3 Officially Released — 2.8 Trillion Parameters, 1M Token Context
China's Moonshot AI officially released its latest model, Kimi K3, on July 16, available on the web and in its app. The model packs 2.8 trillion (2.8T) parameters with a 1 million token context window. The company claims industry-leading performance in coding, agentic tasks, long-form reasoning, visual understanding, and multi-agent orchestration.
The Financial Times had reported in advance that Kimi K3 would become "the largest open AI model out of China," with a parameter count in the 2–3 trillion range. It is expected to narrow the gap with Anthropic's Opus 4.8.
The model had been reported as "launching soon" since last weekend, and it is now finally available for general use. The LocalLLaMA community reacted immediately as well.
Apple Intelligence Wins Chinese Regulatory Approval — Partners with Alibaba and Baidu
Apple's AI assistant, Apple Intelligence, has secured approval to roll out in the Chinese market. Partnerships with Alibaba's Qwen AI and Baidu are now in place, paving the way for AI features to reach iPhone users in mainland China.
The partnership had been rumored since last year, and it has now been formally approved following regulatory review by the Chinese government. China is one of Apple's most critical markets, and the absence of AI features directly undermines its competitiveness, making this approval a major milestone.
Alibaba is expected to provide the LLM foundation, while Baidu will handle search and content moderation.
METI Launches Domestic Physical AI "FRONTia Project"
On July 16, Japan's Ministry of Economy, Trade and Industry (METI) held a kickoff event for the FRONTia Project, which aims to build foundation models for domestically developed physical AI. Centered on Noetra — an AI development company jointly funded by 44 Japanese companies — and the National Institute of Advanced Industrial Science and Technology (AIST), the project targets "real-world-native AI" by 2030.
NVIDIA CEO Jensen Huang took the stage at the event, stressing that "building Japan AI is a must." Minister of Economy, Trade and Industry Ryosei Akazawa also showed up in a leather jacket, signaling a whole-of-nation commitment to the effort. Physical AI — real-world AI such as robotics and autonomous driving — is positioned as a key domain where Japan can leverage its manufacturing strengths.
Sakana AI Integrates Nvidia Nemotron into Fugu — Countering Frontier Models with "Collective Intelligence"
Tokyo-based Sakana AI has integrated NVIDIA's open-source model Nemotron into its orchestrator, Fugu. Fugu is a system that dynamically combines multiple language models depending on the task, validating the approach that "individual open models fall short of frontier models on their own, but become competitive when made to collaborate."
No concrete benchmarks have been announced yet, but the claim is that — as with hybrid autoregressive (AR) and diffusion systems — coordinating multiple models, rather than a winner-takes-all approach, is a promising direction.
AMI Labs CEO LeBrun Rejects the "AGI" and "Superintelligence" Labels
Alexandre LeBrun, CEO of AMI Labs — Yann LeCun's world-model startup — has dismissed the term "superintelligence" that seemingly everyone in the AI industry is chasing. In an interview with TechCrunch, LeBrun explicitly refused to call the company's AI "AGI" or "superintelligence."
AMI Labs is implementing the "world model" approach LeCun advocates, aiming to achieve real-world understanding with an architecture distinct from existing LLMs. Its avoidance of overhyped terminology appears to reflect the careful nature of the company's technical approach.
Lila Sciences — "The Lab Is the Datacenter": An AI-Driven, Autonomous Laboratory
Lila Sciences is building a "lab of the future" where AI autonomously runs experiments around the clock. In an interview with Latent Space, CTO Andy Beam and CSO Rafa Gómez-Bombarelli laid out the company's vision.
Its automated labs feature magnetic levitation transport systems and AI-guided robotics, simultaneously covering biology, chemistry, drug discovery, and materials science. The company says the science-reasoning tokens it has experimentally validated so far exceed 10 trillion. Its core thesis: "Internet data is exhausted, and science is the last untapped data source."
The scale is still being established, but the approach of treating the lab as an infinite token generator points to a noteworthy direction for applying AI to science.
---