This brief covers AI news from 2026-08-29 UTC.
Open-weight models, big money for AI infrastructure, and a landmark FTC ruling shape today’s news.
Tencent open-sources Hy4 preview with 770B parameters
What happened: Tencent released and open-sourced Hy4 preview, a 770B-parameter MoE model with 49B active parameters and a 1M-token context window. It’s available via WorkBuddy, CodeBuddy, Yuanbao, and API through Tencent Cloud TokenHub and OpenRouter.
Why it matters: Developers get a top-tier open model for coding and productivity, with free access for two weeks on WorkBuddy and CodeBuddy.
Z.ai releases GLM-5.3 weights for coding and security
What happened: Z.ai released GLM-5.3 weights, a 756GB model with 256 routed experts and 1M-token context, under a custom license. It scored 28.3 on Terminal-Bench 3.0, up from 4.6 for GLM-5.2.
Why it matters: Developers can self-host a model built for long-running coding and vulnerability hunting, with support for Transformers, vLLM, and SGLang.
a16z raises $1.1B Machine Age fund for AI hardware
What happened: Andreessen Horowitz launched a $1.1B fund focused on AI infrastructure: chips, memory, data centers, and robots. The firm says it’s time to ‘accelerate the physical buildout of AI.’
Why it matters: More capital for hardware startups could ease supply constraints that bottleneck AI development.
Lambda raises $1B in debt to buy Nvidia chips
What happened: Lambda secured $1B in private short-dated debt, arranged by JP Morgan, to buy Nvidia AI chips that it will lease to Microsoft. This follows a $926M loan for GB300 GPUs.
Why it matters: Neoclouds are scaling GPU supply fast, giving developers more rental options for high-end chips.
FTC narrows AI tool provider liability in Rytr vacatur
What happened: The FTC set aside its 2024 consent order against Rytr, establishing a three-part test for the means-and-instrumentalities theory. Providers are liable only if they supply deceptive materials, offer inherently deceptive products, or know of misuse.
Why it matters: Developers of dual-use AI tools face less legal exposure for how users deploy their products.
AMD ROCm 10 delivers 3.3x inference uplift
What happened: AMD released ROCm 10 with ROCm.AI, a new developer experience. Systems with ROCm.AI show an average 3.3x increase in inference performance and 2.4x in training versus ROCm 7.
Why it matters: AMD is closing the software gap, making its GPUs a more viable option for running AI workloads.
Huawei Cloud CodeArts Agent goes GA in Asia Pacific
What happened: Huawei Cloud CodeArts Agent is now commercially available in Asia Pacific, with Basic and Professional editions. It includes Agent Team with 16 specialized agents and 30+ Huawei engineering skills.
Why it matters: Enterprise developers in the region get an agentic coding tool that handles complex, long-duration tasks.