TokenGate is a learned token-pruning module that can be dropped into existing language models. Its authors report a 1.4x to 11.3x reduction in time to first token on long contexts and 1.5x to 3.4x lower peak memory, with no loss of accuracy.
Nvidia's open data engine Flora labels an hour of robot video for 72 cents, against $88 with GPT-6 Astra, with slightly higher quality scores. Built on three fine-tuned Qwen3.5 models, it is far better at spotting errors in robot actions but sometimes less accurate in its descriptions, and Nvidia sees it as part of a push to cover the whole robotics value chain.
A new P-Zero Research benchmark measures how well AI models design and assess experiments on their own. Scores are rising so quickly that Jack Clark now believes self-improving AI could arrive by 2027.
A peer-reviewed study on Z-Library by two Nigerian university librarians cites a TorrentFreak article the site never published. At least ten of its 44 references cannot be found, and the authors acknowledge using two AI tools.
Hugging Face's Carbon-X is an open 14.8-billion-parameter DNA language model built on Qwen3-14B. It reads raw DNA in a chat and answers in English, averages 66.7% across five DNA benchmarks, and is released under Apache 2.0.
The researchers behind OpenMobile-2 argue smartphone agents improve by training on many different apps rather than on more data. They built an open environment with over 110 apps, including 35 replicas of popular ones, and say their model scores highest among open agents on real Android devices.