Meta releases Muse Spark Open: an 816B MoE model that runs on two M5 Ultra Mac Studios
816B total parameters, 42B active per token, with a 1M-token context.
Tip
Huge models are starting to run on desktop machines; how close is “local AI” to ordinary people
Nvidia open-weights Nemotron 3.5 Super VL
A 120B multimodal agent model with 12B active parameters and a 1M-token context.
Tip
Why a chip company releases its own open models
FTC subpoenas OpenAI and Anthropic over AI safety incidents
Following the EU, US regulators are also formally investigating large-model companies.
Tip
With the US and EU acting at once, what exactly counts as an AI “safety incident”
Accenture caps employee Claude spend at $29 a month
Large companies are starting to budget their employees’ AI usage.
Tip
What to do when your company AI budget runs out, and how to get more done with fewer tokens
Top AI models failed a test of inventing new AI research
Models still cannot independently propose and complete new AI research.
Tip
AI is good at execution but not yet at originality; what that means for how we divide work
Alibaba benchmark: the best coding agent produces secure, working code only 25.1% of the time
Much of the code that runs has security problems.
Tip
After AI writes your code, what should a human still check
⚠️ Compiled by an AI assistant from public AI news roundups; original articles are not individually verified. For reference only.