Good morning. The open-weights story keeps gathering force: GLM-5.2 — which we flagged Sunday on release-time alone — now has the benchmarks to back the hype, taking the top spot on Artificial Analysis’s intelligence index. Meanwhile the Fable mess refuses to resolve, the export-control rationale is looking shakier by the day, and Midjourney has pivoted into, of all things, full-body ultrasound machines.
GLM-5.2 takes the open-weights crown. Z.ai’s model now sits at #1 of 92 on the Artificial Analysis Intelligence Index with a score of 51, eleven points above GLM-5.1 and well clear of MiniMax-M3 and DeepSeek V4 Pro at 44. Same 744B/40B architecture, expanded 1M context, unchanged pricing. The accompanying benchmark breakdown highlights an unusually strong non-hallucination rate, and one HN commenter running multi-agent coding evals reported it as the first model they’ve tested that matches or beats Opus 4.6.
GLM-5.2 is built for long context that actually works. Z.ai’s own technical post leans into long-horizon coding as the target use case, with a new IndexShare architecture that cuts per-token FLOPs by 2.9× at 1M context. The MIT license carries no regional restrictions — a detail worth holding onto for the next story. On FrontierSWE and SWE-Marathon it trails only Opus 4.8 while beating GPT-5.5, and the team specifically pitches the 1M window as engineered for agent reliability rather than benchmark bragging.
Anthropic’s export-control fight gets weirder. The Verge reports the administration is using export control rules to bar foreign nationals from accessing Fable 5 and Mythos 5 — a category of regulation that has historically applied to transferable goods, not cloud-based services, with no public legal basis offered. A separate Wired piece covers the parallel demand that Anthropic patch all jailbreaks before reinstatement, which independent security researchers describe as fundamentally not a thing you can do. Tina Nguyen’s column at The Verge reads the whole episode as competing leak campaigns, with both sides floating self-serving versions of who knew what when.
DeepSeek dodges the blacklist. Reuters reports the US has held off on adding DeepSeek to the entity list, even as it flags over 100 other firms as security risks. Worth noting: Z.ai, maker of the just-crowned GLM-5.2, has been on that list since January, and as one HN commenter pointed out, entity listing doesn’t actually block Americans from downloading open weights. The selective enforcement is starting to look less like strategy and more like improvisation.
Local models cross a threshold. Vicki Boykis argues that on a 2022 M2 Mac with 64GB RAM, Gemma 4 and Qwen variants via LM Studio now handle agentic coding at roughly 75% the quality of frontier models — good enough for refactoring, tests, and bootstrapping repos. The HN thread split predictably, with one user claiming Qwen3.6-27B feels like an upgrade over Claude Sonnet 4.6 (“so many strongly-held opinions I did not ever ask it for”) and others insisting the gap to frontier remains large for serious work. The common thread across both camps: API providers should be watching the price ceiling come down.
Midjourney builds an ultrasound machine. CEO David Holz unveiled a full-body ultrasound CT scanner using thousands of transducers in a water-filled ring, developed with Butterfly Network and pitched for deployment in San Francisco spas. The Verge has the details; the Midjourney blog post makes broader claims including “a billion scans a month.” HN reactions ranged from a phased-array specialist saying the underlying beamforming looks plausible to a physician calling it a scam outright, with the VC-funded health data collection angle drawing particular skepticism.
That’s the morning. The Fable standoff isn’t ending, GLM-5.2 makes the case that open weights are a real alternative rather than a backup plan, and Midjourney is apparently a hardware company now. Back tomorrow.