Reflection announced Beam, a mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, aimed at coding, reasoning and agentic work. The company is taking early-access sign-ups while completing red-teaming and evaluations; weights, a technical report and developer materials are planned for later in October. Its performance and inference-efficiency comparisons are company-reported, rather than proof that it will outperform alternatives on your projects. For builders, the useful distinction is between an announced open-weight direction and weights you can download today. Beam belongs on an evaluation watchlist; production plans still need the released artifacts, licensing details and workload-specific tests. For learners, its sparse architecture illustrates why total parameter count alone is a poor shortcut for estimating the computation used on each task.
AI news
Reflection previews Beam; OpenAI introduces text watermarking
OpenAI announced textGrain, which places a statistical watermark in generated wording. API customers worldwide can opt in for selected models, with watermarking off by default. Eligible ChatGPT and Codex outputs in the EU are scheduled to receive it over the following weeks. Detector access initially requires approval for researchers and expert organizations. The educational implication needs care: a watermark is a limited provenance signal, not a measure of human effort or a reliable plagiarism verdict. OpenAI says short, edited or translated text can evade detection, while false positives remain possible. An absent watermark therefore cannot establish human authorship. Educators and publishers should avoid treating a detector result as a complete account of how someone produced a piece of work.