Shanghai AI Lab put Atria Dawn’s weights on Hugging Face days before the paper explaining them: a quiet release of an open 744B agentic model.
Tag: open source
PrismML shrank a 27B model to just three weight values per parameter — and kept 98% of its benchmark performance intact.
AutoArk’s Edge0 runs a 35B model on a 24GB Mac mini by predicting which experts it will need from SSD before it needs them.
DeepSeek published the engineering behind DeepSeek-V4.1-Flash: a 1M-token context held in just 890 bytes of memory per token.