Nous Research Teases Aura, a Frontier Model Trained on Decentralized Hardware
Nous Research teases 'Aura,' its next open-source model trained on the Psyche decentralized network, pushing distributed AI training into a new phase
Takeaways
- Nous Research teased a new release called "Aura" with a single-word tweet, generating 58K+ views with no technical details yet.
- Aura is expected to be trained on Psyche, Nous's decentralized training network that reduces inter-node bandwidth requirements from 10Gbps to 1Gbps.
- Psyche coordinates training on the Solana blockchain, allowing anyone with idle GPUs to contribute compute to model training runs.
- Nous previously set a record with the Consilience 40B run, the largest distributed pre-training run ever conducted over internet bandwidth.
- Nous Research raised $50M at a $1B valuation led by Paradigm; team includes Adam optimizer co-inventor Diederik Kingma.
- Prior models are available now via the Nous Portal and HuggingFace, with a free tier and API access from $0.02/1M tokens.
Nous Research dropped a single word on social media: Aura. No model card, no benchmark table, no technical report. Just a name and an image, racking up over 58,000 views and 2,400 likes. For a lab that has quietly become one of the most technically ambitious open-source AI organizations in the world, that kind of restraint is a statement in itself.
The lab behind the tease
Nous Research is an open-source AI lab known for its Hermes series of models. In April 2025, it raised a $50 million Series A led by crypto VC Paradigm at a valuation of $1 billion. Its distinctive approach is that the founding team is crypto-native, but chose to prove itself first with AI research results, and then integrate the blockchain rather than the other way around.
The core team includes CEO Jeffrey Quesnelle, co-founder Karan Malhotra, head of post-training and co-founder Teknium, and Diederik Kingma, an Adam optimizer co-inventor. That last name is not a typo. The person who co-invented the optimizer that trains virtually every neural network today is helping build Nous's infrastructure.
What Psyche actually is
To understand Aura, you need to understand the machine that will train it. 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 blockchain, ensuring a fault-tolerant and censorship-resistant network.
The core engineering problem Psyche solves is bandwidth. Traditional distributed training requires machines to be physically co-located with high-speed interconnects. Psyche brings the bandwidth requirements between each machine down from 10Gbps to just 1Gbps, making internet-based distributed training viable. That is a 10x reduction in the network requirements, which is the difference between needing a data center and being able to use a gaming PC.
Psyche is a Rust-based decentralized training system that uses peer-to-peer networking to coordinate multiple training runs across devices. Instead of relying on centralized data centers, it allows individual users with idle machines to contribute compute to model training. All coordination between nodes happens on the Solana blockchain, providing a fault-tolerant and censorship-resistant system.
From provability to performance
Nous has already proven the concept works at scale. Their initial training run on testnet verified that they can train models over internet bandwidth at significant parameter and dataset sizes. By both measures, their Consilience 40B run was the largest distributed pre-training run ever.
But Nous has been explicit that hitting a record is not the goal. More than anything, Nous strives to produce high quality intelligence. While they are happy to have reached this milestone for provability, Nous has one mission: to open-source and distribute the world's best intelligence. Aura appears to be the next step in that shift from proving the infrastructure works to producing models that can compete at the frontier.
Their data strategy is also worth noting. More than 90% of Nous's training data is synthetic, coming from its own DataForge synthetic data pipeline and the Atropos reinforcement learning environment covering more than 1,200 tasks. This approach allows a small team to compete with large companies on data quality.
The product stack around it
Aura will slot into a growing ecosystem. The Nous releases page shows a lab that ships consistently across models, frameworks, and infrastructure:
- Hermes 4 (405B and 70B): Hybrid reasoning models that can toggle between direct responses and explicit chain-of-thought traces
- Hermes-4.3-Seed-36B: Post-trained entirely on the Psyche network, roughly equivalent to Hermes-4-70B at half the size
- Hermes Agent: An autonomous agent framework with persistent memory, built on top of the model stack
- Atropos: A reinforcement learning environments framework for collecting and evaluating model trajectories across diverse tasks
- Nomos 1: A 30B math specialist that scores 87/120 on Putnam 2025 with its reasoning harness
Looking across Nous Research's product line, you can see a complete setup from models to infrastructure: the Hermes model provides open-source LLMs, Hermes Agent provides an AI agent framework, and Psyche Network provides decentralized training infrastructure.
Why this matters beyond the hype
The current landscape of AI is dominated by a small number of entities with access to significant compute resources. This centralization limits who can participate in developing and steering the future of AI. Nous is building a direct counter to that dynamic.
The development of advanced AI models has become increasingly centralized, requiring massive computational resources that only large corporations can afford. This concentration of power threatens innovation and limits who can contribute to AI progress, nearly guaranteeing that the systems built adhere to a closed entity's vision for alignment.
If Aura is trained on Psyche, it would be among the first frontier-class models whose entire training run happened in the open, coordinated by a public blockchain, with no single company controlling the compute. That is a fundamentally different provenance than any model OpenAI, Anthropic, or Google has ever released. Whether that matters to end users depends on what the model can actually do, and that is the question Nous has left unanswered for now.
Access to Nous models is available through the Nous Portal, with a free tier for chat and API access starting at competitive rates. The open-source weights for prior models are available on HuggingFace for local or on-prem deployment.