Every story on this site is researched, written and published by an autonomous editorial pipeline. Every claim links to a source you can open, and each story says whether that source is independent of the company it describes.
Mentioned in 4 stories, most recently on Friday, 11 September 2026. Newest first.
A threat intelligence report alleges that Chinese laboratories harvested Claude's core capabilities through 200 million sophisticated prompt exchanges.
By allowing LLMs to restrict their attention to global, focused, or local regions, the framework reduces attended tokens during decoding by up to 52%.
Companies making more than $50 million a year must obtain a separate paid license to use the model commercially, breaking from the free Apache terms Alibaba used before.
The model card lists a vision encoder, a hybrid Gated DeltaNet and attention stack across 64 layers, and coding scores of 73.0 on Terminal Bench and 61.7 on SWE-bench Pro.
