THIS WEEK IN PAPERS
Smaller models, longer memory, cheaper checks
Weekly brief · 4 min read · Aug 4, 2026
Trending on Hugging Face this week
Three papers dominated the trending list this week, and they rhyme more than the titles suggest. Two of them shrink the model and keep the score; the third makes evaluation cheap enough to run on every commit.
The first replaces a 70B dense model with a 12B mixture-of-experts and matches it on reasoning benchmarks at roughly a fifth of the serving cost. The interesting part for a product team is not the benchmark — it is that the cost curve moved enough to change what you can ship in a free tier.
The second extends usable context to 1M tokens without the usual quality collapse past 200K. Retrieval does not go away, but the failure mode changes: you stop losing the middle of the document and start paying for latency instead.
The third is an eval paper. It shows a small judge model, calibrated on 300 human labels, tracking full human review closely enough to gate releases. If it holds up outside the paper, the honest reading is that eval stops being a quarterly project and becomes part of CI.
