Baidu releases open ERNIE 5.5 Flash base model
Baidu released ERNIE 5.5 Flash Base, a sparse mixture-of-experts model that activates only 14B of its parameters per token. The open weights come straight from the end of pre-training, giving developers a clean starting point for further training and domain adaptation. — ernie.baidu.com
MNN-Pocket brings compact models to on-device AI
Alibaba's MNN team shipped MNN-Pocket-0.3B, a small language model family co-designed with its inference stack for on-device deployment, squeezing the quantized build into 173 MB while keeping near-full quality. — huggingface.co
Tracker maps the world's AI security institutes
A new tracker from Accelerate Europe maps the global AI security institute landscape, which 11 countries have now established, with others considering one. It compares mandates, funding, staffing, and access to frontier models, each figure linked to its original source. — lesswrong.com
Answer me with HTML turns replies into readable pages
Answer me with HTML is an agent skill that has the model write a Markdown draft while a bundled CLI renders the finished page, diagrams and all, cutting output tokens to about an eighth of hand-written HTML. — github.com
Heretic gpt-oss-20b gets GGUF builds for local use
A Hugging Face upload offers GGUF quantizations of gpt-oss-20b-heretic, the uncensored variant of the open gpt-oss-20b model, in a range of sizes for running locally. — huggingface.co