NEWSFERENCE
FRI, 15 May 2026 20:32:23
LIVE
$ today --liveF1TodayF2YesterdayF3ArchiveF4About
NEXT SCAN —
← BACK TO TODAY/CLUSTER · NOUS · RESEARCH
CLUSTER · TIER 2
FIRST SEEN 186D AGO
NOUSRESEARCH

Nous Research releases Lighthouse Attention for faster long-context pre-training

Nous Research has open-sourced Lighthouse Attention, a selection-based hierarchical attention mechanism for long-context pre-training that delivers a 1.4–1.7× wall-clock speedup at 98K context and runs ~17× faster than standard attention at 512K context on a single B200. The approach uses a multi-resolution pyramid with top-k cascade selection and requires no custom sparse attention kernel, straight-through estimator, or auxiliary loss.

Sources
1
X mentions
89k ▲
First seen
186Dago
Velocity
+2%/6h
CONTRIBUTING SOURCES
1 ARTICLES
  1. X (Twitter)186D AGO
    x.com/NousResearch/status/2055337939270332862
X DISCOURSE
89k TOTAL · TOP 3
@NousResearch3H · 35.2K
Hermes Cloud agents now open in a browser tab with the exact same interface as Hermes Desktop. https://t.co/4Vw3XJgFwn To launch a personal agent, pick a name and a model in Nous Portal. We host it for you and it runs 24/7 in its own container, so it can handle scheduled jobs and answer you on any channel while your laptop is closed. Chats, Group Chats, plugins and settings in the Webapp work the same way they do when you run Hermes on your own computer. Open your agent from Nous Portal on your phone or any computer, or connect Hermes Desktop to it and continue the same chats there. Plus, Super, and Ultra plans each include one Cloud agent out of the box. Using Hermes from anywhere has never been easier.
@NousResearch23H · 82.8K
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: https://t.co/iAOANQTVZ6
@NousResearch2D · 81.9K
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: https://t.co/iAOANQTVZ6