Issue W46/25: Moonshot AI Launches 1T Kimi K2 Model, Google Unveils TPU v7, and Terminal-Bench 2.0 Debuts - November 10, 2025
Last week, Moonshot AI launched Kimi K2 Thinking, a 1T parameter open-weights model claiming state-of-the-art agentic performance, while Google announced its 10x more powerful TPU v7 chip, and the Terminal-Bench 2.0 benchmark was released for more rigorous agent evaluation.
General News
Soumith Chintala, who led PyTorch since its inception, announced his departure from Meta after 11 years. Amid ongoing debates about compute financing, Sam Altman clarified that his vision involves a broader U.S. reindustrialization effort for the AI supply chain, not government loan guarantees for OpenAI. In robotics, XPeng announced its IRON gynoid will enter mass production in late 2026. On the consumer front, Perplexity will become the default AI in Snapchat chat starting January 2026.
Google DeepMind
Google announced its 7th-gen TPU, "Ironwood," will be generally available in the coming weeks, promising a 10x peak performance improvement over TPU v5p and will be used to train and serve Gemini. DeepMind released IMO-Bench, a suite of benchmarks for advanced mathematical reasoning. On the product side, Gemini's Deep Research feature can now draw context from Workspace apps like Gmail and Drive, and Google AI Studio launched a managed RAG tool called File Search.
Foundation Labs & Model Updates
Moonshot AI released Kimi K2 Thinking, a 1T parameter open-weights Mixture-of-Experts model with ~32B active parameters, a 256K context window, and native INT4 precision. The model claims state-of-the-art results on agentic benchmarks like HLE and BrowseComp, and independent analysis has noted its strong agentic and coding performance. It is already available via vLLM and Ollama, and has been shown running on Apple Silicon via MLX. Elsewhere, OpenAI announced capacity and rate-limit improvements for its Codex models, and Meta released EdgeTAM, a real-time segment tracker that is ~22x faster than SAM2.
AI Developer Topics
Developer tooling for agents is maturing, with VS Code introducing a unified "Agent sessions" view and Anthropic publishing guides on efficient code execution with MCP. LangChain expanded its ecosystem with Deep Agents for JavaScript/TypeScript, while llama.cpp launched a new polished WebUI for easier local model interaction.
Research News
The Terminal-Bench 2.0 benchmark was released with tougher tasks and cloud container support via the new Harbor framework. Perplexity shared research on custom kernels for trillion-parameter MoEs optimized for standard cloud infrastructure. Additionally, new research introduced DreamGym for using synthetic environments in RL, and Cambrian-S for improving spatial awareness in video models.
Others Topics
XPENG unveiled its IRON humanoid robot, sparking discussion about its human-like gait and internal mechanics. In creative AI, Coca-Cola released another AI-generated Christmas advertisement, continuing its shift toward automated production despite mixed public reactions. In Switzerland, a major supermarket is selling a cookie box with an AI-generated design that features a reindeer with five legs.