This brief covers AI news from 2026-11-13 UTC.
A leadership shock at the chipmaker every AI lab depends on, plus a new Google default model and a 145B model that runs on a laptop.
Jensen Huang resigns as NVIDIA CEO, CFO Kress steps in
What happened: NVIDIA’s board announced on Nov 12 that Jensen Huang resigned as president, CEO and board member, effective immediately. CFO Colette Kress is interim CEO while a successor is sought. Reports link the departure to alleged affairs; the allegations are unconfirmed.
Why it matters: NVIDIA supplies most LLM training GPUs, so teams planning capacity or pricing around its roadmap now face a leadership question with no named successor.
Google ships Gemini 4 Hydrogen as the new default model
What happened: Google released Gemini 4 Hydrogen, succeeding Gemini 3.8 Flash and sitting alongside Carbon and Argon. It has a 1M token context at $1.50 per million input and $7.50 per million output tokens, and is the default in AI Mode and Search.
Why it matters: The cheaper Flash tier now carries a 1M context, so long-document and multi-part query work can move off the pricier reasoning models.
Edge0 open-sources a 145B model that runs on a laptop
What happened: Edge0 released Edge0-145B and Edge0-120B with an open-source inference framework and deployment tools. Internal tests report 29GB peak memory, 1, 100+ tokens/s prefill and 80+ tokens/s decode, with 10B parameters active per step.
Why it matters: If the numbers hold, coding and agent tasks can run locally with no per-token API bill and data that never leaves the machine.
Mirendil raises $1B Series A led by Kleiner Perkins
What happened: Mirendil, founded by about 20 researchers from Anthropic, xAI, Google DeepMind and OpenAI, announced a $1B Series A led by Kleiner Perkins and Clearlake, with Google participating. It builds the Mirendil-Base coding model, the Fern 0.1 agentic model and Lor, a platform for delegating work to agents in private beta.
Why it matters: A new lab with a billion dollars and a coding model plus an agent platform is a direct competitor for teams choosing a model provider.
Anthropic ships Claude for Manufacturing with 17 skills
What happened: Anthropic released the Claude for Manufacturing plugin with 17 skills, including quoting, failed-inspection investigation, breakdown diagnosis and supplier sourcing, plus four connectors including Autodesk Fusion and Infor.
Why it matters: Factory and hardware teams get packaged agent skills wired to the systems they already run, instead of building tool integrations from scratch.
Mercor drops OWE-Bench for long-horizon engineering tasks
What happened: Mercor released OWE-Bench, 26 long-horizon engineering and research tasks graded on unseen cases with continuous scoring. Claude Fable 5.2 leads at 54.6% using an average of 211 million tokens per attempt; GPT-6 Astra follows at 44.2%.
Why it matters: Continuous scoring shows how far an agent gets on real engineering work, which is more useful for picking a model than pass-or-fail coding tests.
Thinking Machines posts Inkling 2 Small weights under Apache 2.0
What happened: Thinking Machines published the weights of Inkling 2 Small on Hugging Face under an Apache 2.0 license, the release it had promised when it launched the model. The multimodal model has 594B total parameters and 17B active per token.
Why it matters: Apache 2.0 permits commercial use, so teams can self-host and fine-tune the model without a vendor agreement.