Nous Research
Decentralized AI startup developing open-source, human-centric language models and tools.
Website: https://nousresearch.com/
Cover Block
PUBLIC
| Field | Value |
|---|---|
| Name | Nous Research |
| Tagline | Decentralized AI startup developing open-source, human-centric language models and tools |
| Headquarters | New York City, United States |
| Founded | 2023 |
| Stage | Series A |
| Business Model | Open Source / Commercial |
| Industry | Deeptech (AI / Machine Learning) |
| Technology Type | LLM architecture, distributed training, agentic tooling |
| Geography | North America |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2): Jeffrey Quesnelle, Karan Malhotra |
| Funding Label | $50M+ |
| Total Disclosed | ~$70,000,000 |
Links
PUBLIC
- Website: https://nousresearch.com/
- LinkedIn: https://www.linkedin.com/company/nousresearch
- Hermes Agent: https://hermes-agent.nousresearch.com
Executive Summary
PUBLIC
Nous Research is a New York-based applied research group betting that the next generation of frontier-grade language models can be trained, owned, and operated by an open community rather than by a small number of hyperscalers [Crunchbase]. Founded in 2023 by Jeffrey Quesnelle and Karan Malhotra, the company grew out of a volunteer collective of independent researchers before formalizing into a venture-backed entity, and its early credibility came from technical contributions such as the YaRN paper on context-window extension co-authored by Quesnelle and chief scientist Bowen Peng [LinkedIn]. The company's most distinctive bet is DisTrO, a method for distributed model training across the public internet, and its successor system Psyche, which in December 2024 reportedly completed a test training run of a 15 billion-parameter model across 11,000 steps on a global network [aibase.com]; [Andreessen Horowitz]. Commercially, Nous ships open-weights Hermes models and a Hermes Agent framework that learns user projects and works across multiple model providers [GitHub]. The capital base is now significant for an open-source-first lab: roughly $70 million in disclosed funding, anchored by a $50 million Series A led by Paradigm in April 2025 at a reported $1 billion token valuation, with prior seed participation from OSS Capital, Distributed Global, North Island Ventures and Delphi Digital [Fortune, April 2025]; [Crunchbase]. Over the next 12 to 18 months, the watch items are whether Psyche's distributed training stack produces a model that is competitive on standard benchmarks rather than only on novelty, whether the token-linked structure underpinning the Series A translates into durable economic alignment with model users, and whether the team can scale operationally beyond its founding cohort.
Data Accuracy: GREEN -- Confirmed by Fortune, Crunchbase, LinkedIn, and a16z.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series A |
| Business Model | Open Source / Commercial |
| Industry / Vertical | Deeptech, AI / ML |
| Technology Type | Open-weights LLMs, distributed training, agents |
| Geography | North America (NYC) |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | ~$70M disclosed; $1B reported valuation [Fortune, April 2025] |
Company Overview
PUBLIC
Nous Research began as an informal collective of independent AI researchers and model fine-tuners who congregated around open-source LLM work in 2023, and was incorporated the same year with headquarters in New York City [Crunchbase]; [Tracxn]. According to materials published by the company and by LinkedIn, the organization "originated as a volunteer project" and converted dedicated contributors into full-time employees after its first institutional funding round [LinkedIn]. Co-founder Jeffrey Quesnelle serves as chief technology officer and co-founder Karan Malhotra holds the title of director; chief scientist Bowen Peng joined from the same research community [jeffq.com]; [LinkedIn]; [ZoomInfo].
The company's public milestones cluster tightly. A first disclosed seed of $5.2 million led by Distributed Global was reported in January 2024 [CypherHunter, Jan 2024]; [The Block]. A larger $20 million seed led by OSS Capital followed in June 2024 [Crunchbase]. In December 2024, Nous published results from a 15 billion-parameter test training run on its Psyche network, which it framed as a stability proof for distributed training across heterogeneous hardware [aibase.com]. In January 2025 the DisTrO-enabled training code for Psyche was released publicly on GitHub [X/Twitter, Jan 2025]. The company's most consequential capital event came in April 2025, when Paradigm led a $50 million Series A at a reported $1 billion valuation tied to a future token, marking one of the larger crypto-AI crossover rounds of the year [Fortune, April 2025].
Nous's positioning as "a leader in the American open source AI movement, training world-class open source language models" is taken directly from its own site [Nous Research]. PitchBook classifies its primary industry as Software Development Applications, while Tracxn and Crunchbase emphasize the human-centric language model framing [PitchBook]; [Tracxn]; [Crunchbase].
Data Accuracy: GREEN -- Confirmed by Crunchbase, PitchBook, Tracxn, and Fortune.
Product and Technology
MIXED
Nous's product surface today spans three layers: open-weights language models (the Hermes family), a distributed training stack (DisTrO and Psyche), and an agentic application layer (Hermes Agent). The company describes its applied research focus as "LLM architecture, Data Synthesis and Local Inference" on its LinkedIn page [PUBLIC] [LinkedIn], and earlier technical work, notably the YaRN paper on efficient context-window extension co-authored by Bowen Peng and Jeffrey Quesnelle, established initial credibility within the open-source LLM community [PUBLIC] [LinkedIn].
The most differentiated piece of the stack is DisTrO, a distributed training methodology that, per a16z's podcast with the team, is intended to allow geographically dispersed contributors to co-train large models over the public internet rather than inside a single data center [PUBLIC] [Andreessen Horowitz]. Psyche extends DisTrO and a related component called DeMo into what the company describes as "open infrastructure for decentralizing AI training across underutilized hardware" [PUBLIC] [Nous Research]. The December 2024 test, in which Psyche reportedly completed 11,000 training steps on a 15 billion-parameter model spanning a global node set, is the strongest public datapoint on technical viability so far [PUBLIC] [aibase.com]. The training code was published on GitHub in January 2025 [PRIVATE-adjacent confidence: ORANGE] [X/Twitter, Jan 2025]. Fortune's reporting on the Series A notes that Solana is used as a coordination layer in the training process, which is the link between the AI work and the token-denominated valuation [PUBLIC] [Fortune, April 2025].
On the application side, Hermes Agent is an open-source agent framework, hosted on GitHub and a dedicated subdomain, that learns from user projects, accumulates skills, and is designed to be model-provider agnostic [PUBLIC] [GitHub]; [hermes-agent.nousresearch.com]. The company has also referenced an upcoming AI orchestration product called Nous-Forge in its LinkedIn description [PUBLIC] [LinkedIn], though no public release details are available at the time of writing.
Data Accuracy: YELLOW -- Confirmed by a16z and company sources; independent benchmark validation of Psyche's training quality is not yet public.
Market Research and Opportunity
PUBLIC
Decentralized model training sits at the intersection of three of the most heavily capitalized technology markets of the decade: foundation models, AI compute infrastructure, and crypto-native coordination networks. Named third-party TAM estimates specific to decentralized AI training are not present in the cited research, so this section uses adjacent public reference points and the qualitative framing offered by Fortune and a16z.
The demand-side argument for distributed, open-source training rests on two observations that recur across the cited coverage. First, frontier-model training is increasingly capital-rationed: Fortune's reporting on the Paradigm round explicitly frames Nous as a counter to a market "dominated" by closed labs such as OpenAI, with DeepSeek's emergence cited as evidence that competitive open models can be trained for far less than the conventional wisdom suggested [Fortune, April 2025]. Second, large pools of underutilized GPU capacity exist outside hyperscaler data centers, in gaming rigs, smaller cloud providers, and crypto mining facilities, and Psyche is positioned to recruit that capacity into a coordinated training fabric [Nous Research]; [Andreessen Horowitz].
Adjacent and substitute markets matter to the thesis. The most direct substitute is centralized open-weights training by well-capitalized labs (Meta's Llama line, Mistral, DeepSeek, Alibaba's Qwen), which can release weights without needing distributed coordination. The most direct adjacency is the broader "decentralized AI" category that crypto investors have funded in parallel, including Bittensor, Gensyn, and Prime Intellect, each of which targets a slightly different slice of training, inference, or model-marketplace coordination. Regulatory forces cut both ways: open-weights distribution is subject to ongoing scrutiny in both the EU AI Act implementation and US executive guidance, while the token-linked capital structure exposes Nous to the evolving US treatment of crypto-native fundraising.
| Reference point | Value | Source |
|---|---|---|
| Series A round size | $50,000,000 | [Fortune, April 2025] |
| Reported token valuation | $1,000,000,000 | [Fortune, April 2025] |
| Total disclosed funding | ~$70,000,000 | [Tracxn] |
| Psyche test model size | 15B parameters, 11,000 steps | [aibase.com] |
The takeaway from these datapoints is narrow but real: capital availability and a published technical proof point both exist, which puts Nous in a small group of decentralized-training projects with the resources to attempt a competitive open model rather than only a research demo.
Data Accuracy: YELLOW -- Sizing is qualitative; only round-level and Psyche test figures are independently citable.
Competitive Landscape
MIXED
Nous competes simultaneously against centralized open-weights labs that ship strong models without distributed training, and against crypto-native AI projects that pursue distributed training without yet shipping comparably strong models. The structured facts do not name specific competitors, so the analysis below is written as prose and grounded in the categories surfaced by the cited Fortune and a16z coverage rather than in a head-to-head table.
In the centralized open-weights segment, the practical reference points are Meta's Llama series, Mistral, DeepSeek, and Alibaba's Qwen. These groups have shown that open-weights releases can reach or approach frontier benchmarks when backed by either a hyperscaler's compute budget or, in DeepSeek's case, by efficient training methods on a more modest cluster [PUBLIC] [Fortune, April 2025]. For users who only care about weights and license, these incumbents are direct substitutes for Hermes, and they release on a faster cadence than a distributed network can currently match. Nous's defensible edge here is narrative and community, not raw scale: it has cultivated mindshare among independent fine-tuners and agent builders, and Hermes models have become a recognizable brand within that audience.
In the decentralized-AI segment, the relevant peers include Bittensor (incentivized model and subnet marketplace), Gensyn (verifiable distributed training), and Prime Intellect (distributed training with a focus on collaborative runs). Nous's edge versus these peers is that it is, first and foremost, a model lab that also runs a network, rather than a network looking for models to train. The DisTrO and Psyche stack is being validated against Nous's own training objectives, which shortens the feedback loop between infrastructure work and model output [PUBLIC] [Andreessen Horowitz]; [Nous Research]. The exposure is symmetric: if a peer such as Prime Intellect ships a competitive distributed-trained model first, the "first credible decentralized frontier model" narrative that underwrites the $1 billion valuation could migrate elsewhere [PUBLIC] [Fortune, April 2025].
Where Nous is most exposed is in raw compute economics against a determined hyperscaler-backed open lab. A Meta or an Alibaba can absorb the cost of a single training run that a distributed network can only match by aggregating thousands of contributors, and the coordination overhead of distributed training is a real, not rhetorical, tax. Nous does not own a distribution channel comparable to Hugging Face or to a hyperscaler's model garden, which means downstream reach depends on third parties.
The most plausible 18-month scenario is a bifurcation. Winner-if: Psyche completes a publicly verifiable training run of a model that is competitive with a same-generation Llama or Qwen release, at which point Nous becomes the reference implementation for community-trained frontier models and the token economics start to attract serious GPU contribution. Loser-if: distributed training quality lags centralized open releases by more than one model generation through 2026, in which case Nous risks being valued primarily as a crypto-native research brand rather than as a frontier model producer.
Data Accuracy: YELLOW -- Competitor set inferred from category coverage in cited sources; no head-to-head benchmarks are public.
Opportunity
PUBLIC
If Nous executes, the prize is to be the default open coordination layer for community-trained frontier models, a position with no clear incumbent today.
The headline opportunity. The single largest outcome reachable from Nous's current position is becoming the canonical open-source counterpart to closed frontier labs: the place where a competitive, openly licensed model is trained in public, on contributor hardware, with a token-aligned economic layer. Fortune's framing of the Paradigm round explicitly positions Nous as a credible answer to "a market dominated by giants like OpenAI," and cites DeepSeek's cost-efficient training as evidence that the absolute compute gap is narrower than previously assumed [Fortune, April 2025]. The December 2024 Psyche test of a 15 billion-parameter run across 11,000 steps is the most concrete public evidence that the underlying coordination problem is tractable at non-trivial scale [aibase.com]. Together, those two points make the headline outcome reachable rather than purely aspirational.
Growth scenarios.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Reference open lab | Nous ships a Hermes model trained end-to-end on Psyche that ranks within striking distance of Llama or Qwen on standard benchmarks | A successful >50B parameter Psyche training run completes publicly | Psyche has already cleared a 15B/11,000-step stability test [aibase.com] |
| Agent distribution | Hermes Agent becomes a default open agent runtime for developers building model-agnostic workflows | Adoption by a major open-source developer ecosystem or model gateway | Hermes Agent is shipping today as model-provider agnostic [GitHub]; [hermes-agent.nousresearch.com] |
| Token-coordinated GPU network | Psyche's token economics attract a large pool of independent GPU contributors, creating a self-reinforcing training network | Token launch tied to the reported $1B valuation activates contributor incentives | Paradigm-led $50M Series A was structured around a token valuation [Fortune, April 2025] |
What compounding looks like. The flywheel Nous is implicitly building has three loops. First, every additional successful training run on Psyche increases confidence that contributor GPUs will be productively used, which should attract more compute to the next run. Second, every Hermes model that gains traction in the open-source community pulls fine-tuners and agent developers into the Nous ecosystem, deepening the data and feedback loop that informs the next model. Third, if and when a token activates, contributor compute and model usage become economically linked rather than purely reputational, which historically is what has separated durable crypto networks from temporary ones. Early evidence that loop one is starting comes from the public release of Psyche's training code in January 2025 [X/Twitter, Jan 2025], and that loop two is starting comes from the Hermes Agent release [GitHub].
The size of the win. Public comparables for an open frontier lab with a token-linked economic layer are scarce, but two reference points frame the upside. Mistral, a centralized open-weights peer with no token, was reported across 2024 press at valuations in the multi-billion-dollar range, illustrating that open-weights labs can support large equity valuations on their own. Separately, Nous's own Series A was struck at a reported $1 billion token valuation [Fortune, April 2025]. If the reference open lab scenario plays out and Psyche becomes the default decentralized training fabric, the company's combined equity-and-token enterprise value could plausibly compound several multiples beyond entry (scenario, not a forecast). The downside counterweight to this upside lives in the private half of this report.
Data Accuracy: YELLOW -- Upside framing is scenario-based; only round, valuation, and Psyche test figures are independently citable.
Sources
PUBLIC
[Fortune, April 2025] Exclusive: Crypto VC giant Paradigm makes $50 million bet on decentralized AI startup Nous Research at $1 billion token valuation | https://fortune.com/crypto/2025/04/25/paradigm-nous-research-crypto-ai-venture-capital-deepseek-openai-blockchain/
[Crunchbase] Nous Research - Crunchbase Company Profile & Funding | https://www.crunchbase.com/organization/nous-research
[Crunchbase] Series A - Nous Research - Crunchbase Funding Round Profile | https://www.crunchbase.com/funding_round/nous-research-series-a--0c5e9fee
[Crunchbase] Seed Round - Nous Research - Crunchbase Funding Round Profile | https://www.crunchbase.com/funding_round/nous-research-seed--a857b650
[PitchBook] Nous Research 2026 Company Profile: Valuation, Funding & Investors | https://pitchbook.com/profiles/company/572004-37
[Tracxn] Nous Research - 2026 Company Profile, Funding & Competitors | https://tracxn.com/d/companies/nous-research/__sxOeTQ0bR0asJ45fh7NztgITONmEs-3WvnsCDc8GeE8
[Nous Research] NOUS RESEARCH - Open Source AI | https://nousresearch.com/
[LinkedIn] Nous Research company page | https://www.linkedin.com/company/nousresearch
[LinkedIn] Karan Malhotra - Director - Nous Research | https://www.linkedin.com/in/karan-s-malhotra/
[Andreessen Horowitz] DisTrO and the Quest for Community-Trained AI Models | https://a16z.com/podcast/distro-and-the-quest-for-community-trained-ai-models/
[Jeffrey Quesnelle] Homepage of Jeffrey Quesnelle | https://jeffq.com/
[Startup Intros] Nous Research: Funding, Team & Investors | https://startupintros.com/orgs/nous-research
[Into the Bytecode] Jeffrey Quesnelle on Nous Research, large language models, and the human mind | https://open.spotify.com/episode/5dBYAQ6HtC29Mu8VWlYoM6
[aibase.com] Psyche 15B parameter distributed training test coverage | https://aibase.com
[GitHub] Hermes Agent repository | https://github.com
[hermes-agent.nousresearch.com] Hermes Agent product page | https://hermes-agent.nousresearch.com
Articles about Nous Research
- Nous Research Wants Every Idle GPU on the Internet Training the Next Open Model — With $70M from Paradigm and OSS Capital, the New York lab is betting decentralized training can keep open-source AI competitive.