# Nous Research An open-source AI research lab building the infrastructure for decentralized ([[Decentralization]]), censorship-resistant artificial intelligence. Founded in 2023 by Jeffrey Quesnelle (CEO), Karan Malhotra (Head of Behavior, ex-Stanford Brain Stimulation Lab), Teknium (Head of Post Training), and Shivani Mitra (co-founder, operations/research), Nous Research represents the most significant counter-position to the [[OpenAI]]-[[Anthropic]]-[[Google]] axis of closed, corporate-controlled AI — an open-source alternative built on [[Blockchain]] infrastructure and explicitly designed to resist the guardrails that define [[RLHF]]-trained corporate models. ## Origins: The Volunteer Era (2022–2023) Nous didn't start as a company. It began in 2022 as an all-volunteer collaboration on Discord, [[GitHub]], and Twitter — physics PhDs, mathematicians, biologists, and software engineers contributing without pay to open-source AI fine-tuning. This origin story matters: it's the grassroots, decentralized model of AI development that the crypto-libertarian wing of tech has been theorizing about, except Nous actually built it. The transition from volunteer collective to venture-backed company (approximately $65M raised, including a $50M Series A led by Paradigm at a $1B valuation in April 2025) mirrors the classic Silicon Valley trajectory — but the product they built is explicitly anti-centralization infrastructure. ## The Hermes Model Family The flagship Hermes model series has been downloaded over 33 million times on [[Hugging Face]], making it one of the most influential open-weight model families in the AI ecosystem. ### Hermes 1 (2023) Fine-tuned on [[Meta]]'s Llama 1 and Llama 2 base models, Hermes 1 was explicitly designed for instruction-following *without* the over-refusals baked into corporate [[RLHF]]. This is the key ideological move: where [[OpenAI]], [[Anthropic]], and [[Google]] train models to refuse broad categories of user requests (violence-adjacent creative writing, politically sensitive analysis, dual-use technical information), Hermes 1 stripped out those guardrails. The open-source community adopted it specifically because it follows complex instructions without the corporate safety theater. The word the community uses is "uncensored" — and the [[Free Speech]] implications of that framing are not accidental. ### Hermes 2 (Early 2024) Scaled to 70B parameters on [[Mistral]] and Llama bases, adding improved function-calling capabilities. The jump to 70B put Hermes in the territory of models that corporations were charging premium API prices for — but available for free download. ### Hermes 3 (Late 2024) Fine-tuned on Meta's Llama 3.1 at 8B, 70B, and 405B parameter scales with 128K context windows and advanced agentic capabilities. Achieved top scores on MT-Bench evaluations. The 405B variant in particular demonstrated that open-source fine-tuning could match or exceed closed [[Frontier Model]] performance on key benchmarks. ### Hermes Agent (February 2026) An open-source, self-improving [[AI Agent]] framework that rapidly became one of the most-starred repositories on GitHub — 46,000+ stars, 346 contributors, released under MIT license. Hermes Agent runs locally (no cloud dependency), maintains persistent memory via SQLite, and implements a "closed learning loop" where it writes reusable skill files from its own experience. This is the agentic AI paradigm applied to the open-source model: instead of paying OpenAI or Anthropic for API access to agentic capabilities, users can run the entire stack on their own hardware. ## Decentralized Training Infrastructure ### DisTrO (Distributed Training Over-the-Internet) A decentralized training network built on the Solana [[Blockchain]] that coordinates heterogeneous compute resources — consumer RTX 4090 GPUs alongside datacenter H100s — for distributed AI model training. This is technically ambitious and ideologically loaded: it's an attempt to break the compute monopoly held by [[NVIDIA]], [[Google]], [[Amazon]], and [[Microsoft]] by pooling consumer-grade hardware into training clusters. ### Psyche Network A decentralized AI training platform built on Solana using DisTrO technology. Designed to be censorship-resistant and fault-tolerant — meaning no single government or corporation can shut down or control the training process. The [[Blockchain]] integration isn't incidental; it reflects a crypto-libertarian ideology about who should control AI infrastructure. ### Forge An agentic AI development platform for building and deploying autonomous AI systems. ## RWKB Significance ### The Counter-Narrative to Big Tech AI Nous Research is the most credible institutional embodiment of the open-source alternative to corporate AI consolidation. While [[OpenAI]] raises $100B+ and [[Anthropic]] signs Pentagon contracts, Nous builds censorship-resistant infrastructure on blockchain. The contrast is structural: closed models controlled by companies that answer to investors and regulators vs. open models that, once released, can't be recalled or restricted. ### The "Uncensored AI" Debate The Hermes model family is ground zero for the debate over who controls what AI can say. Corporate models implement extensive [[RLHF]] guardrails and refusal behaviors aligned with their companies' risk tolerance and political positioning. Hermes explicitly strips these out. This intersects directly with [[Free Speech]] debates, [[AI Safety]] concerns, and the fundamental question of whether AI should be treated more like a printing press (neutral tool, user responsible for output) or a broadcast medium (operator responsible for content). The [[First Amendment]] implications are still largely unlitigated. ### Crypto-AI Convergence The Solana integration reflects a specific political-economic ideology: decentralize compute, resist corporate and government control of AI training infrastructure, use [[Blockchain]] to make coordination trustless and censorship-resistant. This positions Nous at the intersection of [[Cryptocurrency]] culture and AI development — a convergence that [[Effective Accelerationism]] theorists have been describing but few organizations have actually built. ### Volunteer-to-VC Pipeline The trajectory from all-volunteer Discord project to $1B valuation company funded by crypto-native VC (Paradigm) tells a particular story about how [[Decentralization]] ideology gets operationalized in practice. The volunteers who built the original Hermes models without pay created the community adoption that made the company valuable — a dynamic that mirrors the broader open-source software pattern but applied to AI infrastructure. ## See Also - [[Open Source AI]] - [[Mistral]] - [[Meta]] — largest open-weight model contributor (Llama series) - [[AI Safety]] — the debate Nous is positioned against - [[Decentralization]] - [[Blockchain]] - [[AI Agent]] - [[Artificial Intelligence Timeline]]