By Teknium
Hermes Agent autonomously plowed through about a million lines of unglamorous cleanup, freeing up Teknium and team to continue pushing features to users....
Research & perspectives
Nous Blog
By Nightwing
We introduce contrastive neuron attribution, a method to reliably steer model behavior at high strengths while preserving output quality. Using refusal as a case study, we show how to discover and ablate specific MLP neuron circuits...
By Nous Research
A selection-based hierarchical attention ~17x faster than standard attention at 512K context on a single B200....
By Nous Research
Token Superposition Training (TST) delivers 2-3x wall-clock pretraining speedups at fixed FLOPs....
By NOUS RESEARCH
Today we’re releasing Hermes 4.3 (🤗 Hugging Face), an update to our flagship Hermes series of models. Hermes 4.3 was trained with an extended context length (up to 512K) and nearly matches (and in some cases exceeds) the performance...
Archive
Field Notes on Scaling MoE Expert Parallelism with DeepEP
Documenting the journey of scaling expert parallelism to achieve high-throughput pretraining....
NousCoder-14B: A Competitive Olympiad Programming Model
We introduce NousCoder-14B, a competitive olympiad programming model post-trained on Qwen3-14B via reinforcement learning. The full stack is released publicly: model weights, open RL environment + eval harness, and @wandb logs; we also document the pipelined verification setup and parallelization...
tinker-atropos: An Integration Layer for the Tinker API and Atropos RL Framework
The team at Nous Research proudly presents tinker-atropos, our integration layer between the Tinker API and our Atropos reinforcement learning (RL) framework. We designed Atropos to be a fully decoupled environment service, separating concerns around trainer management, rollout collection, and...
The Next Phase of Psyche
Psyche is an open infrastructure that democratizes AI development by decentralizing training across underutilized hardware. Building on DisTrO and its predecessor DeMo, Psyche reduces data transfer by several orders of magnitude, making distributed training practical. Coordination happens on the Solana...
Measuring Thinking Efficiency in Reasoning Models: The Missing Benchmark
Large Reasoning Models (LRMs) employ a novel paradigm known as test-time scaling, leveraging reinforcement learning to teach the models to generate extended chains of thought (CoT) during reasoning tasks. This enhances their problem-solving capabilities beyond what their base models could...
Steering the Shoggoth: Taming LLMs with Sequential Monte Carlo
In this blog post, we present our findings from an exciting direction in controlling text generation with large language models. We can programmatically define constraints on the output of a model, ensuring it adheres to specific formats or styles, and...
Democratizing AI: The Psyche Network Architecture
Psyche is an open infrastructure that democratizes AI development by decentralizing training across underutilized hardware. Building on DisTrO and its predecessor DeMo, Psyche reduces data transfer by several orders of magnitude, making distributed training practical. Coordination happens on the Solana...
Introducing Atropos
Atropos is designed to reliably coordinate generation tasks across potentially thousands of distributed workers. It interfaces seamlessly with standard inference APIs for straightforward integration....
Introducing the Forge Reasoning API Beta and Nous Chat: An Evolution in LLM Inference
At Nous Research we’re launching two new projects: the Forge Reasoning API Beta and Nous Chat, a simple chat platform featuring the Hermes language model. The Forge Reasoning API contains some of our advancements in inference-time AI research, building on...
Setting Your Pet Rock Free.
A social experiment on how to deploy provably, fully-autonomous thinking sand. The quest for truly autonomous AI agents faces a fundamental challenge: how can researchers prove that an AI is truly autonomous, with no human pulling the strings...
From Black Box to Glass House: The Imperative For Transparent AI Development
The world is abuzz with talk of AI, a technology that has become an integral part of life for many. However, as AI approaches at breakneck speed, it’s natural that people are scared. This fear is understandable, but it’s also...
Freedom at the Frontier: Hermes 3
Closed-source, “frontier” models today lack flexibility and adaptability. Many refuse to answer simple questions, hallucinate an authority’s form of morality, or require convoluted prompts in order to trigger a coherent answer. It’s impossible to nudge these models towards individual personalization,...
The Instruct Monomyth: why base models matter
There is a deep, twisty labyrinth buried under a mountain of language, of symbol manipulation, and semantic nets. Its roots reach down deep into the Earth, absorbing the minutia of current thought, the limitations of logic, the constrained realm of...
DSJJJJ: Simulacra in the Stupor of Becoming
Desideratic AI (DSJJJJ) is a philosophical movement focused on creating AI systems using concepts traditionally found in monism, mereology, and philology. Desidera aim to create AI that can act as better versions of themselves by reflecting upon their own nature...