The Big Story
Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%. According to a report from MarkTechPost, Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B. This groundbreaking achievement has significant implications for the field of natural language processing.
The new model boasts impressive results on Nous's 803-prompt eval, where propaganda answers fell from a concerning 22% to a remarkably low 0.9%. Moreover, Hermes Bench stayed flat at 65.60%, indicating a notable increase in model accuracy and effectiveness. The development of Arete-27B marks a crucial step forward in the quest for more reliable and trustworthy AI models.
The significance of this breakthrough cannot be overstated. As AI becomes increasingly integral to our daily lives, ensuring that these systems are free from biases and propaganda is essential. The success of Arete-27B demonstrates the potential for fine-tuning existing models to achieve remarkable improvements in terms of accuracy and reliability.
What Shipped
Google AI has released a new plugin for spec-first coding agents called Elephant-Goldfish Model, which makes drafting specs easy before running 5 zero-context review stages.
A config-verified comparison of how 10 open MoE models route tokens was published by MarkTechPost. The post highlights the total and active parameters, expert counts, top-k, shared experts, gating functions, and load-balancing methods used in each model.
Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B, which significantly cuts propaganda from 22% to 0.9% on Nous's 803-prompt eval, with Hermes Bench staying flat at 65.60%. This marks a crucial step forward in the quest for more reliable and trustworthy AI models.
A comprehensive coding guide for NVIDIA's Dory 1.0 4B looped language model was published by MarkTechPost. The guide includes building the pinned llama.cpp branch, analyzing loop-by-loop compute costs, testing memory caching, and running parallelized inference on GPU accelerators.
From the Labs
The Elephant-Goldfish Model plugin from Google AI makes drafting specs easy for coding agents by running 5 zero-context review stages before any code.
MarkTechPost published a config-verified comparison of how 10 open MoE models route tokens. The post highlights the total and active parameters, expert counts, top-k, shared experts, gating functions, and load-balancing methods used in each model.
A comprehensive coding guide for NVIDIA's Dory 1.0 4B looped language model was published by MarkTechPost. The guide includes building the pinned llama.cpp branch, analyzing loop-by-loop compute costs, testing memory caching, and running parallelized inference on GPU accelerators.
Other Notable News
A new guide has been published on MarkTechPost, explaining MoE routing strategies and how different models use them. The post covers seven methods, including Mixtral's top-2 softmax and DeepSeek-V4's hash routing.
NVIDIA Dory 4B has received a comprehensive coding guide from MarkTechPost, providing step-by-step instructions for building the pinned llama.cpp branch, analyzing loop-by-loop compute costs, testing memory caching, and running parallelized inference on GPU accelerators.
A comparison of how different MoE models route tokens was published by MarkTechPost. The post highlights the total and active parameters, expert counts, top-k, shared experts, gating functions, and load-balancing methods used in each model.
Elon Musk's data centers are facing uncertainty after the midterm elections, according to a report from TechCrunch. The article examines what this could mean for the tech industry as a whole.
A memo reportedly claimed that China recruited a female technology executive to spy on Nvidia's CEO, according to an article from TechCrunch. This alleged recruitment was said to have occurred earlier this year.
The Take
The Elephant-Goldfish Model plugin from Google AI has brought coding agents to a new level by making them draft specs and then running five zero-context review stages before any code is written. This move has significant implications for the way we approach software development, as it prioritizes quality control and precision.
This shift in paradigm is also reflected in the realm of MoE (Mixture of Experts) models, where routing strategies have become a critical aspect of model architecture. The likes of DeepSeek-V4's hash routing, Mixtral's top-2 softmax, and Qwen's dynamic routing are just a few examples of the innovative approaches being explored.
In related news, Nous Research has released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B, which has achieved impressive results in reducing propaganda answers from 22% to 0.9%. This development highlights the importance of fine-tuning models for specific tasks and applications.
In a separate but equally significant development, NVIDIA's Dory 4B model has been made more accessible through a comprehensive coding guide that walks users through building the pinned llama.cpp branch, analyzing loop-by-loop compute costs, testing memory caching, and running parallel processing.