Subscribe
Share
Share
Embed
China's open-weights strategy is both a product and a policy problem
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Thursday, June eighteenth.
In today's briefing we see Zhipu AI's GLM-five point two claiming the top spot on open-weights benchmarks under MIT licensing, the US holding off DeepSeek enforcement as a diplomatic card rather than a security trigger, Anthropic opening its Seoul office with major chaebol partnerships, and Midjourney pivoting into medical imaging with an ultrasonic CT scanner.
First up - Today in the big model news;
Gemini
Gemini three point five Pro is expected within the next two weeks. Sundar Pichai's timeline at Google I/O puts the launch in the June twenty-two to June twenty-six window. The model targets a two-million-token context window and Deep Think reasoning, positioning itself as a direct competitor to Anthropic's Mythos and GPT five point five. For product teams deciding between model providers, the practical move is to wait for independent benchmarks rather than finalizing your selection now, because once a new frontier model releases, the previous comparison set becomes obsolete.
Anthropic - Claude
Anthropic opened its Seoul office on June seventeenth, paired with a memorandum of understanding with South Korea's Ministry of Science and ICT and enterprise rollouts at major chaebols: NAVER running Claude Code across the entire engineering organization, plus Samsung SDS, LG CNS, and Nexon. Korea is already a top-twelve Claude.ai usage market, and the chaebol structure creates multiplicative leverage. A single enterprise relationship at Samsung or LG cascades through dozens of subsidiary companies simultaneously. For product leaders deciding where to concentrate international presence, the model is to go deep in markets where demand is already established and enterprise leverage is multiplicative, because chaebol architectures mean one major contract propagates across dozens of subsidiary relationships.
China's open-weights strategy
Zhipu AI's GLM-five point two claimed the highest score on Artificial Analysis's Intelligence Index, delivered on its June fourteenth promise of MIT-licensed open-weights availability. As of June seventeenth, the model is publicly available: seven hundred and forty-four billion total parameters with forty billion active in a mixture-of-experts configuration, scoring fifty-one on the Intelligence Index against competitors at forty-four. Pricing is one point four dollars per million input tokens and four point four dollars per million output tokens, well below frontier closed-source rates. One structural signal: GLM-five point two generates forty-three thousand output tokens per task on average versus twenty-four to thirty-seven thousand for competitors, a compounding difference at scale. For AI PMs building enterprise deployments, frontier-quality open-weights models at substantially reduced costs now clear a meaningful threshold, because what's available under MIT licensing has crossed into genuine production capability.
This connects to the second story: the US interagency security committee approved adding DeepSeek to the Commerce Entity List along with over one hundred other entities, but Commerce Under Secretary Jeffrey Kessler delayed execution to avoid escalating trade tensions with China. The security case is substantive. Anthropic and OpenAI have separately warned Congress that DeepSeek ran active campaigns to extract model capabilities via API distillation, and there are documented attempts to acquire restricted US chips through Southeast Asian shell companies. The delay is not about insufficient evidence; the committee signed off. It reflects treating blacklist authority as diplomatic currency. For product teams evaluating Chinese AI services, the distinction between MIT-licensed open-weights models and active API services under security committee review is material, because already-distributed MIT-licensed weights operate under different enforcement and deployment timelines than active services.
In other news:
Midjourney Medical announced a full-body ultrasonic CT scanner using eighty-nine hundred and sixty transducers, producing MRI-grade imaging without radiation, completing a full-body scan in sixty seconds for a few dollars each. That's roughly ten times cheaper than a traditional MRI. The company licensed the underlying ultrasound-on-chip technology from Butterfly Network in November twenty twenty-five and plans to open a spa location in San Francisco's Union Square in twenty twenty-seven as its first commercial deployment, with a stated goal of fifty thousand scanners worldwide by twenty thirty-one. For teams shipping vision AI into vertical markets, the strategic opening is medical imaging, where visual intelligence interpretation commands the highest dollar value, because multimodal visual AI trained at scale is exactly the tool needed to reconstruct diagnostic scans from ultrasonic transducer data.
That's the briefing. Have a great day.