Arete-27B builds on Qwen3.8-27B. Its Apache 2.0 weight files total about 54.71 GB across 18 files before any lower-precision conversion.
Nous reports training on about 52.8 million tokens, units of information a model processes, over two passes. It chose Inkling’s answers as its teacher and aimed to reinforce what it calls Western values.
On its own 803-prompt test, Nous reports that answers adopting Chinese state framing fell from 22% to 0.9%. Its fact score rose from 2.32 to 2.98 out of 5, with the two models tested at matching settings.
A judge model, a model grading another model’s output, scored the answers. The released test says the same judge also screened which questions were answerable, and its questions are English-only.
The training-data card warns that most accepted examples still carry a factual-error tag. Its displayed accepted-row count differs from the article’s training-example count.
Caveat: The results measure the developers’ chosen questions and grading rules, not political neutrality or accuracy across viewpoints.
(Techmeme · Nous Research’s model card, test and dataset documentation)
An agent is a model running a multi-step task with tools. The author contrasts moving memory between products with letting several tools read and write the same shared address.
The proposed record includes a scope, key, author and expiry. The author also suggests common read and write operations and tests for preserving unknown fields.
The account compares existing export formats and memory stores, and acknowledges that shared memories can carry malicious instructions. Attribution and signatures can help trace a record, but do not prevent poisoning.
Caveat: This is a vendor’s proposal, not an adopted standard or a demonstrated interoperability test.
(HN 59 points · 52 comments · LoreKit proposal; author builds the product)
Inner Mongolia data-center visit → LOW_UTILITY: the accessible overview describes infrastructure rather than a directly usable tool, and detailed methods are subscriber-only.