Ternary-Bonsai-2-27B: A 27B Model in 6GB
A 27B model squeezed into 5.95GB at 1.72 bits per weight, claiming 98.2% of full-precision performance on its own benchmark suite - and what that compression actually costs.
A 27B model squeezed into 5.95GB at 1.72 bits per weight, claiming 98.2% of full-precision performance on its own benchmark suite - and what that compression actually costs.
Alibaba built transparency into the image format itself rather than removing backgrounds afterwards - then shipped it under a research licence instead of the Apache 2.0 its predecessors used.
DeepSeek-V4.1-Flash is bigger than its predecessor and cheaper to run. The savings come from cutting KV cache to 890 bytes per token - about a quarter of the previous model.
A 35-author arXiv paper calls graph engineering a new paradigm; LangChain says it is the latest name for what LangGraph has done since 2023. Both have a point.
A developer turned off Claude Code automatic memory and called it his best upgrade - because inferred memory is an invisible layer between a decision and its reason.
An open-source memory layer where memories fade on a 7-day half-life unless retrieval keeps them alive - and an unrelated arXiv paper that independently landed on the same idea.
What this site is, and what to expect.