In a predawn vote on July 1 the Senate adopted Marsha Blackburn's amendment striking the AI section from the reconciliation bill 99-1, with Thom Tillis the only member voting to keep it. Authored in the Senate by Ted Cruz, the provision would have barred states from enforcing AI laws for a decade and, in its funding-linked form, gated BEAD broadband money on states standing down.

It follows the House's 215-214 passage in May, opposition from 17 Republican governors and 40 state attorneys general, and a Cruz-Blackburn compromise that cut the moratorium to five years and survived less than a day. The Senate then passed the bill 51-50 on the vice president's tiebreak, and the moratorium is not in the law signed on July 4.

Zuckerberg told staff Monday that every AI team at Meta — foundations, products, FAIR — now reports into Meta Superintelligence Labs, led by ex-Scale AI CEO Alexandr Wang as Chief AI Officer with ex-GitHub CEO Nat Friedman over products and applied research. The memo names eleven researchers hired from OpenAI, Anthropic and Google DeepMind, keeps Llama 4.1 and 4.2 on the roadmap, and starts a small new lab on frontier models.

It closes a month in which Meta invested $14.3 billion for a 49% stake in Scale AI and, by multiple accounts, dangled packages worth up to $100 million at individual researchers. OpenAI's chief research officer told staff in a Slack memo that leadership was recalibrating compensation while the company went dark for a week to recharge, warning that Meta would use the pause to pressure people into fast decisions.

The same Monday, Huawei released open weights for its Pangu line: a 7B dense model and Pangu Pro MoE 72B with 16 billion active parameters, plus Ascend-optimized inference code. Huawei describes Pro MoE as the strongest sub-100B model it has published, built on a Mixture-of-Grouped-Experts router designed to even out expert load in distributed training.

The framing is ecosystem, not charity: weights and serving code that make Huawei's accelerators a workable target without CUDA. Huawei said the 7B weights would follow shortly; the Pro MoE weights and its inference stack shipped on day one, and the company invited outside developers and researchers to download and feed back.

Baidu published weights and inference code for all ten ERNIE 4.5 variants on Monday, including an MoE flagship with 424 billion total parameters and 47 billion active per token, down to a 0.3B dense model, all under Apache 2.0 with 128K context on the language variants. Baidu had shipped the same family as a closed API in March.

The company reports ERNIE-4.5-300B-A47B-Base beating DeepSeek-V3-671B-A37B-Base on 22 of 28 benchmarks, and a 21B-A3B variant competitive with Qwen3-30B-A3B at roughly 30% fewer total parameters. Training and serving run on Baidu's own PaddlePaddle stack, with 4-bit and 2-bit quantization and the ERNIEKit fine-tuning toolkit open-sourced alongside.