Model Release · Front Page

Reflection ships Beam, a 501B open agent model

The 501-billion-parameter model keeps only 23B active at a time, built for frontier reasoning, coding, and scientific workloads — and it arrives, reportedly, without distillation.

Reflection's long-awaited first release lands with a deliberately efficient pitch. Beam carries 501 billion total parameters but activates only 23 billion at once, a sparse mixture-of-experts design aimed at keeping long-horizon agentic work cheap enough to run in production. The company positions the model for frontier reasoning, coding, and scientific workloads, joining a field-wide shift from models that simply answer toward models that act.

Early reaction was sharp and immediate. Researchers pointed out that Beam appears to ship without any distillation from rival frontier models — a claim that, if it holds, would make it a rare clean signal for how far open-weight training can go on its own. Others framed it as the latest American lab to release a top model while still trailing its Chinese counterparts on the strongest benchmarks.

Policy · Provenance

OpenAI extends provenance to text for the EU

Content verification moves beyond images and audio as regulators tighten the AI Act.

OpenAI is expanding its content-provenance tooling to include text, a step it says responds to EU regulatory requirements under the AI Act. The company is candid about the limits: current text-watermarking technology remains unreliable under editing and paraphrasing, and it is framing the rollout as a first step rather than a solved problem. Existing tools already let users check whether an image or audio file was created with OpenAI models.

Research · Reinforcement Learning

Coding harnesses become RL environments

A striking result for the open-source RL community: coding harnesses including Claude Code, Codex, Hermes, Pi, and opencode were turned into reinforcement-learning environments without touching the harnesses or the training code. Any open model, any task set, fully open source. With the same model and the same weights, the setup reaches 62% on Mini-SWE-Agent — evidence that environment engineering, not model tweaking, is becoming the lever that matters.

Industry · Competition

Meta and Microsoft tighten internal use of Claude

The Information reports that both Meta and Microsoft are steering employees away from Anthropic's Claude and toward their own models and tools. Inside Meta, the number of employees using Claude Code fell from roughly 60,000 at the start of the year to about 30,000 recently, a drop partly attributed to spring layoffs but also to an explicit push toward in-house alternatives.

Scale · Infrastructure

1.3 billion sandboxes in four weeks

A back-of-the-envelope estimate puts the frontier in perspective: 1.3 billion sandboxes over four weeks works out to roughly 48.46 million per day. At DeepSeek, a single DSec-scale unit handles about 3 million, which means reproducing that throughput would demand roughly sixteen such units — or, by one reading, around 656 GB300s each. It is a big project, and a rare hint at how reinforcement learning is actually run at the frontier.