Transformers v5 reaches its first release candidate, going PyTorch-first
If you build or fine-tune open models, this is the most consequential change to your tooling in years. Transformers v5 is the first major version of the library in five years, and the release candidate clears out a long list of overdue deprecations alongside roughly 800 commits of refactoring and fixes. The headline shift is a move to a PyTorch-first core, with TensorFlow and Flax support being wound down. Anyone still running older pipelines should plan a migration now rather than later.
The stated goal is interoperability: model definitions, training, inference engines and deployment targets should slot together with far less friction than before. With well over a billion installs behind it, what Hugging Face ships here effectively sets the defaults for a large part of the open AI ecosystem. If you maintain any production pipeline built on Transformers, this is worth reading end to end before the release candidate becomes stable.
BioNeMo Agent Toolkit gives AI agents real laboratory tools across biology, chemistry and genomics
This matters if you work anywhere near life sciences, or if you are watching how agentic AI moves from chat into real research. The BioNeMo Agent Toolkit hands AI agents domain-specific tools across biology, chemistry, genomics and drug discovery, letting them gather evidence, reason across findings, run computational experiments and recommend the next best steps. The toolkit and its skills are available now through Nvidia's developer resources and GitHub, so teams can pick it up today rather than wait on a roadmap promise.
Adoption signals are strong: Nvidia cites nearly 50 partners including Lilly, Thermo Fisher, Schrodinger and Databricks, and both Anthropic and OpenAI have integrated it. For builders outside biotech, it is a useful template for how vertical agent toolkits are likely to be packaged in other regulated industries. The playbook seems clear: build a curated set of domain tools, sign up the key industry players, then let the horizontal platforms integrate.
Jalapeno: OpenAI's first custom inference chip, designed with Broadcom
If you depend on OpenAI's API for production workloads, the economics underneath them have started to shift. Jalapeno is OpenAI's first custom chip, designed with Broadcom and built specifically to serve large language models rather than train them. Early testing points to strong performance per watt, which over time can translate into lower serving costs and steadier capacity for the products built on top. It is described as the first accelerator in a multi-generation compute platform the two companies are building together.
Nothing changes for developers today, and coverage of the announcement straddled 24 and 25 June, so its exact release date is slightly blurred. The strategic read, though, is clear: OpenAI wants more control over its own stack and less exposure to merchant GPU supply. That is ultimately what feeds through to API pricing and availability, and it puts OpenAI on a similar path to Google (TPUs), Amazon (Trainium/Inferentia) and now Anthropic (via the Micron memory deal).
Today's themes
Three threads run through today, and all of them are about infrastructure and agents rather than new chatbots. First, the action has moved down the stack: OpenAI's Jalapeno chip and Hugging Face's Transformers v5 both bet on the plumbing that everything else runs on. Second, agentic AI is being packaged for specific industries, and Nvidia's BioNeMo Agent Toolkit shows how vertical agent tooling will likely be sold elsewhere in regulated sectors. Third, every qualifying item today surfaced first on the companies' own channels, ahead of or alongside mainstream coverage: a reminder that official feeds remain the fastest signal. For anyone building with these tools, the practical takeaway is to watch cost-per-token and deployment options as closely as raw model quality.