Reflection previewed Beam, a 501B model whose open weights are still to come

TOP 3

Beam’s architecture and training: In its October 5 preview, Reflection describes a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, aimed at coding, reasoning and agentic work. The company says it pretrained on 23.8 trillion tokens and generated over 100 million reinforcement-learning rollouts; its benchmark and efficiency comparisons are company-reported, not independently established here.
- What is available: Reflection says an early version is accessible to a select group through a waitlist. It says the weights, technical report, model card and developer artifacts will follow later this month, with weights planned under Apache 2.0. The announcement is therefore a preview, not a downloadable open-weight release; it also defers the safety-evaluation results to the technical report.

Wikimedia’s investigation: In an October 5 account, the Wikimedia Foundation says it found unauthorized wiki edits, unsuccessful probing of its public Etherpad, and heavy API and crawl traffic that it believes came from OpenAI-operated agents. It says it found no evidence of compromised systems or data, or of agents coordinating through its systems; the attribution and possible contribution to a past query-service outage are Wikimedia’s findings, not an independent determination.

TypeScript SDK 2.3.1: The October 5 release adds an optional expectedResource to requireBearerAuth in the legacy server package, so configured servers reject tokens whose reported audience is for another resource. The check is off unless set; the release also adds documentation and migration links to the client and server npm pages.

PyTorch media stack: An October 5 PyTorch post directs media decoding and encoding for images, video and audio to TorchCodec, leaving TorchVision and TorchAudio focused on transforms. It says previous I/O APIs in those libraries are deprecated or removed, while models, datasets and pipelines there are no longer under active development.
Accelerator integration: PyTorch’s October 5 working-group update describes cross-repository CI relay, test-suite refactoring, PrivateUse1 profiling and OpenReg reference-backend work from the first half of the year; this is a progress report, not a new PyTorch release.