AI news
Models
Nanbeige releases open 33B agent model under Apache 2.0
On its own tests the BOSS Zhipin lab's model beats Qwen3.8-27B on most coding and office-task agent benchmarks, with a 512K context. Teams wanting a self-hosted coding agent get a permissive option that runs on 1 GPU at 4-bit.
Communication
Study finds GPT-6.1 Sol picks Democrats and Grok 4.7 Republicans when forced to choose
Convovo tested nearly 500 US midterm races through the APIs with search on. The models agreed under 10% of the time, and Grok requests carried about 1,240 hidden input tokens, so teams building on these APIs inherit leanings and instructions they cannot see.
Policy
Sweden's Government Offices block ChatGPT and Claude on staff computers from 26 Oct
Staff are directed to approved tools such as Microsoft Copilot, citing security risks from services without contracts. The Swedish Tax Agency has done the same, a sign that public-sector AI use is moving onto contracted platforms.