This Week in AI: Kimi K3, Inkling, and Japan's National AI Factory
Moonshot AI pushed open weights past the 3T mark with Kimi K3, Thinking Machines released its first multimodal Inkling model, and Japan committed to a Vera Rubin national AI factory while agent harnesses went open source.
This week tilted toward open frontier models and the infrastructure around them: trillion-parameter weights from China, a new multimodal base from Mira Murati's lab, Japan staking national compute on physical AI, and a wave of agent harnesses shipping or opening their code.
Models
Moonshot AI releases Kimi K3 — Moonshot unveiled a 2.8-trillion-parameter model with native vision and a one-million-token context window, calling it the first open 3T-class system. Weights are promised by July 27, with API and product access live now across Kimi.com, Kimi Work, and Kimi Code.
Thinking Machines Lab open-sources Inkling — Mira Murati's startup released its first in-house model: a 975B-parameter MoE that reasons over text, images, and audio under an Apache 2.0 license. Inkling targets customization via the Tinker fine-tuning platform rather than raw benchmark leadership.
NVIDIA ships Nemotron 3 Embed — NVIDIA published a family of open embedding models, with the 8B variant ranking first on the RTEB multilingual leaderboard. Smaller 1B variants and an NVFP4 Blackwell-optimized build aim at production RAG and agentic retrieval at lower latency.
Agents & Tools
xAI open-sources Grok Build — SpaceXAI published the Rust agent harness behind its coding CLI under Apache 2.0, exposing the agent loop, tool layer, TUI, and extension system for skills, plugins, hooks, and MCP servers. Teams can compile it locally and point it at their own inference endpoints.
Microsoft Agent Skills for Python reaches stable release — The Agent Framework's skills API graduated from preview, letting Python agents discover domain expertise on demand through progressive disclosure. Human-in-the-loop approval gates script execution and resource loading by default.
LM Studio launches Bionic agent app — LM Studio introduced a separate agent-focused app for coding, research, and document work with local or cloud open models. It adds voice input, sandboxed file operations, automatic checkpoints, and optional access to frontier open weights through LM Studio Secure Cloud.
Policy & Infrastructure
Demis Hassabis proposes a FINRA-style frontier AI regulator — The Google DeepMind CEO argued for an industry-funded, government-supervised standards body that would review frontier models up to 30 days before release. The plan responds to criticism of opaque U.S. government holds on recent Anthropic and OpenAI launches.
Japan and NVIDIA launch national Vera Rubin AI factory — Noetra Corp., backed by METI and a consortium including SoftBank and Sony, will build a 140-megawatt facility with 27,500 Rubin GPUs and 13,750 Vera CPUs. The site targets open multimodal foundation models for robotics, digital twins, and physical AI under Japan's FRONTia program.
The through-line is control without closing the frontier: bigger open weights, inspectable agent harnesses, and governments treating compute as industrial policy. Builders get more choice in models and tooling; policymakers are racing to define who tests what before it ships.