Reflection ships Beam, a 501B open-weight model aimed at China's frontier labs

Good morning 👋 An Nvidia-backed lab says its new model matches a Chinese frontier release on a quarter of the compute. Its own benchmark table tells a slightly different story.

In today's issue:

• 🔭 Reflection ships Beam, a 501B open-weight model aimed at China's frontier labs
• 🧠 Claude goes GovCloud GA, ChatGPT gets an ads format
• 🔬 Base-model reasoning, cheaper agent memory, and an idea-theft detector
• 🏛️ OpenAI's EU watermark compliance plan, with a catch
• 🛠️ Stress-test Beam against your own inference bill
• 🔗 SignSplit's debut round, Moonshot's IPO runway, AI vs. breast cancer

🔭 THE ONE THING

🔓 Reflection ships Beam, a 501B open-weight model aimed at China's frontier labs

Reflection AI, the Nvidia-backed startup that's raised roughly $4.7B total, shipped Beam yesterday: 501B total parameters, 23B active in a MoE, trained on 23.8T tokens, context stretched to 1M. The pitch is parity with Z.ai's GLM-5.2 at 3 to 4x less inference compute, a claim TechCrunch notes hasn't been independently verified. Reflection's own benchmark table undercuts the framing some: Beam trails GLM-5.2 on HLE no-tools, 36.2% to 40.5%. Weights don't actually land until later this month, so for now it's a waitlist and a blog post. Promising shape for the open-weight side of the US-China model race, just not proof yet.

🧠 MODELS & RELEASES

• 🔐 Anthropic Claude Opus 5.5 and Sonnet 5.5 are now GA in AWS GovCloud, hooked into Claude Code for regulated and ITAR work. Agencies that couldn't touch Claude for compliance reasons just lost that excuse. Supercharge regulated workloads with Claude Code and Amazon Bedrock
• 📢 OpenAI ChatGPT is getting a visual ad format, with attribution partners and brand-suitability tools bolted on for advertisers. The thing that answers your question is about to also pitch you a product. Building advertising for the way people use AI

🔬 RESEARCH HIGHLIGHTS

• Base models already knew how to reason. Researchers found that forcing a base model's answer to open with the right starting tokens gets it close to the performance of the same model after RL fine-tuning for reasoning, no reward model required. The capability was sitting in the pretraining data all along. If that holds up at scale, a slice of the reasoning-RL pipeline is spending real compute to teach a model something a prompt template could've triggered for free. arXiv

• MemPilot builds agent memory on request, not up front. Most agent memory systems preprocess and store everything a session touches, which burns compute and still manages to drop details that matter later. MemPilot curates memory on demand instead, pulling and structuring only what a given query needs, across text and other modalities. Not a smarter agent. A cheaper one, which for anyone running agents at volume is the metric that actually shows up on the invoice. arXiv

• IdeaLens checks who had the idea, not who typed it. Standard AI detectors look at the words on the page. IdeaLens flags AI-originated ideas even when a human wrote every sentence, aimed at the provenance question policies are starting to care about more than authorship. Academic integrity offices will want this before any product does. arXiv

🏛️ POLICY & REGULATION

• OpenAI spelled out how it's handling the EU AI Act's text-provenance mandate: watermark generated text, build a detector for it, then hand that detector to researchers first, not the public. The compliance approach binds OpenAI under the Act's content-marking rule, and the staggered rollout is the real tell. They're not ready to let outsiders stress-test the detector yet.

🛠️ TRY THIS

Pressure-test Beam against your current inference bill before you migrate anything

1. Pull the new aws-ai-ml skill into your coding agent (Claude Code, Kiro, or Codex all support it via the Agent Toolkit for AWS).

2. Point it at Beam's open weights and ask for a SageMaker inference benchmark, not a deployment. You want numbers first.

3. Run that benchmark against whatever's serving your traffic now, latency and cost per 1K tokens.

4. If Beam wins on both, start a staged swap. If it only wins on one, you've got a pricing lever for your current vendor, not a reason to move.

Prompt: Using the aws-ai-ml skill, generate a SageMaker inference benchmark comparing Beam-7B against our current production endpoint on latency and cost per 1K tokens, then flag which config is cheaper at our actual traffic volume.

🔗 QUICK LINKS

• SignSplit Launched with a $400M round from W Group, pricing the data-licensing startup at $1B before it has a marquee AI lab signed as a customer. SiliconANGLE
• Moonshot AI Closed its last pre-IPO round at a $50B valuation, up from $31.5B over the summer, and is lining up a Hong Kong listing for early 2027. Bloomberg
• NVIDIA Its blog rounds up startups layering AI onto mammography and pathology to catch breast cancer earlier. NVIDIA

Beam's weights land later this month. We'll have real numbers then, not just Reflection's.

Pradeep Perugu