Some part of me is excited for the coming catastrophe(s), because it'll end the frustrating last ~12mo of "those Anthropic jerks just want regulation so they can corner the market on artificial minds wrought from sand!". Another part of me knows that;
A) Rapid epistemic decompression of this scale will probably lead to some truly nasty, illogical sentiments, not to mention other pathologies (deep depression, compulsive anxiety, etc.). Maybe some will reasses their beliefs after some sober reflection, but a minotiry, I expect.
B) Trading hundreds, thousands, millions, or billions of lives so that I feel a little less terrified and alone isn't exactly compelling utilitarian calculus...
Anyway, to the tech: another impressive jump for Alibaba, catching Moonshot overall with some interesting niche improvements and a massive, gaping flaw in the poor TerminalBench scores. I'm probably just biased as a SWE, though. I can't help but feel like Ziphu will be the ones to really blow our socks off next, given how long it's been since GLM-5.3[1] and the vague rumours about the next one (5.5...?).
The Chinese labs' leaps to catch up to Anthropic & OAI from the strange, long-forgotten past of September 2026 were clearly due to a tripartite distillation campaign, so expansive and rushed that they accidently leaked tons of CCP state secrets.[2] Has anyone heard any rumours about if that's been continuing after the crackdown? Anthropic has been pretty damn aggressive with catching distillation, so my personal guess is that this is just honest improvements.
...which would mean that at least 2 of the 4 Chinese labs are ~3-5 months from the RSI possibilities Anthropic is trying to sound the alarm about today, if that. Jesus fucking christ. We need to rally around international protections ASAP, cause something tells me at least one of the four won't quite so ready to "pace" themselves in the current climate.
---
[1]: https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
[2]: https://www.anthropic.com/threat-intelligence-report-septemb...
This is not an open-weights modell. You have to wait for the smaller models
"Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best."
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
dan_thinks•35m ago
Looking forward to the "Flash" and 27B versions.