River AI (legally River AI Inc.) is an American artificial intelligence company based in Palo Alto, California, founded in 2026 by Igor Babuschkin, a co-founder of xAI. The company says its mission is to create "personal AI owned and shaped by each individual,"[25] and it describes a plan to rebuild the AI stack from training infrastructure up to personal hardware.[1][13] Its first product is the River API, a service for fine-tuning, reinforcement learning and serving open-weight large language models, sold as the managed River Cloud or as River Cluster software on a customer's own GPUs.[12] River came out of stealth on June 10, 2026,[1][3] and on August 11, 2026 it announced $1.1 billion in funding across a Series Seed and a Series A, led by General Catalyst and AMP PBC with strategic investment from Nvidia and AMD Ventures.[2][3] The company did not disclose a valuation.[4][5]

Field	Detail
Legal name	River AI Inc.[12][26]
Type	Private company
Incorporated	April 20, 2026, in Nevada[5][6]
Public launch	June 10, 2026[1][7]
Co-founder and CEO	Igor Babuschkin[1][2]
Headquarters	Palo Alto, California; second office in Austin, Texas[5][24]
Products	River API, River Cloud, River Cluster, River Inference, River Console, River Recipes[12][14][19]
Funding	$1.1 billion (Series Seed and Series A, announced August 11, 2026)[2]
Lead investors	General Catalyst, AMP PBC[2]
Valuation	Not disclosed[4][5]
Website	river.ai

History

Background

Babuschkin spent the decade before River at Google DeepMind, OpenAI and xAI.[1] River's funding announcement says he worked on generative modeling and reinforcement learning at DeepMind and led large-scale training efforts at OpenAI before co-founding xAI with Elon Musk in 2023.[2][8] At DeepMind he was part of the team behind AlphaStar, the StarCraft II system, and in River's launch post he wrote that he had started the AlphaCode project.[1][8] He left xAI on August 13, 2025, saying he would start Babuschkin Ventures, a venture firm meant to support AI safety research and back startups.[8]

Fundraising talks and incorporation

River AI was incorporated in Nevada on April 20, 2026, according to documents reviewed by Forbes.[6] On May 14, 2026, Forbes reported, citing several people familiar with the deal, that Babuschkin was in talks to raise up to $1 billion in initial financing for the company at a valuation of up to $5 billion, with General Catalyst in talks to lead. The same sources said Babuschkin would put up to $100 million of his own money into the company, and that the amount being raised could still change. Babuschkin and General Catalyst did not respond to Forbes's requests for comment.[6]

At that point it was not clear what River would work on. Forbes placed it in a growing wave of so-called "neolabs," startups founded by prominent AI researchers that raise very large sums to pursue frontier, long-term research instead of immediately building and selling products, and quoted an investor who described the company as having been only an idea on a napkin.[6]

Launch (June 2026)

Babuschkin announced River on X on June 10, 2026, writing that it was "a new AI company with the mission to build AI systems that are owned and shaped by you."[7] A launch essay on the company blog, signed by Babuschkin as co-founder and CEO, framed the company around one question: "if machines can do the work, what is left for people?" He argued that replacement of human workers was not inevitable, and that agents which learn from their users and remain under their control could lead somewhere different. He described the River API as "the first piece," sitting at the training layer, and said the next layer would be "personalization and continual learning for agents." He also wrote that the company was keeping its team small on purpose.[1]

The API launched as a "v0.1 preview." Its launch page offered LoRA-based fine-tuning and reinforcement learning on open-source models "from 35B to 1T parameters," starting with Qwen3.6 35B, Qwen3.5 397B, Kimi K2.6 and GLM 5.1, with per-token prices for prompt, completion and training tokens.[11]

Bloomberg covered the launch. According to summaries of its report published by 36Kr and the Russian outlet 3DNews, the founding team included twin brothers Ievgen and Dima Soboliev, former xAI engineers who are River co-founders responsible for engineering; Vincent Stark, who worked on product safety at xAI until December 2025; and Lily Lim, who handled legal matters at xAI (36Kr described her as xAI's former general counsel).[9][10] River's own announcements describe the founding team as having experience at xAI and Tesla.[2]

$1.1 billion funding (August 2026)

On August 11, 2026, River announced $1.1 billion in funding across its Series Seed and Series A. General Catalyst and AMP PBC led the financing, Nvidia and AMD Ventures made strategic investments, and Y Combinator and Temasek also took part.[2][31] TechCrunch described AMP PBC as an AI-focused investment firm founded in 2026 by Anjney Midha, a former general partner at Andreessen Horowitz.[3] The split between the two rounds was not disclosed.[2]

The amount was above the up to $1 billion that Forbes had reported in May.[5] River did not announce a valuation, and Reuters reported that the company declined to comment on how much the funding valued it at.[4][5] AI Weekly, a news aggregator, attributed a valuation of roughly $5 billion to The New York Times, which interviewed Babuschkin for the announcement; River has not confirmed that figure.[28][31] Dealroom, in a summary drawing on the same New York Times report, said that Babuschkin wanted River to build servers that would let individuals and businesses run open-source AI on their own hardware, keeping private data away from River or any other company, and that River had about 20 employees and was still months away from releasing its open-source technologies.[27]

The funding announcement made several claims about the API: that an enterprise could complete a complex reinforcement learning run in 15 to 20 minutes without an infrastructure team, at "two to four times the cost savings relative to closed-source alternatives," and that billing metered on tokens removes the cost of idle GPU capacity.[2] The Next Web noted that these performance claims were the company's own and had not been independently tested.[5]

Date	Event	Amount	Investors	Valuation
May 14, 2026	Talks reported by Forbes	Up to $1 billion (reported target)	General Catalyst reported in talks to lead; Babuschkin reportedly to invest up to $100 million	Up to $5 billion (reported target)[6]
August 11, 2026	Series Seed and Series A announced	$1.1 billion	General Catalyst and AMP PBC (leads); Nvidia and AMD Ventures (strategic); Y Combinator; Temasek	Not disclosed[2][4]

Product expansion (June to October 2026)

After the launch, River added models and features to the API and published research on using it. In October 2026 it launched a separate inference product and released its first open-source training recipe.[14][19]

Date	Release
June 10, 2026	River API v0.1 preview launches with Qwen3.6 35B, Qwen3.5 397B, Kimi K2.6 and GLM 5.1[11]
June 22, 2026	GLM 5.1 correctness fixes and speculative decoding[14]
June 29, 2026	Dropless mixture-of-experts training for the GLM, Kimi and Qwen families[14]
July 6, 2026	GLM 5.2 serving and training; Teams support with shared-wallet credits[14]
August 21, 2026	DeepSeek V4 Flash and Nemotron 3.5 Lightning 30B added[14]
September 3, 2026	GLM-5.3-Flash added[14]
October 6, 2026	River Recipes and a text-to-SQL study published[19][20]
October 7, 2026	Training metrics and projects added to the River Console[14]
October 11, 2026	River Inference launches with DeepSeek V4.1 Flash and GLM 5.3 Flash[14][18]
October 15, 2026	Day-zero support for Nvidia's Nemotron 3.5 Super VL[14][23]

Products and technology

River API

The River API is a service for training and serving language models from Python. Its documentation describes three ways to change a model: supervised fine-tuning on examples, reinforcement learning from rewards, and distillation from stronger teacher models. The user controls the learning loop in Python, and River runs the training and inference.[17]

Training uses low-rank adapters (LoRA).[2][11] The examples on River's website show a client library, river_client, in which a user opens a session, creates a model from an open-weight base with a LoRA configuration (the rank, and whether to train the attention projections, the MLP and expert layers, and the unembedding layer), and then calls low-level operations such as sample, forward_backward, optim_step and save_weights. The user supplies the data, the reward function and the advantages, and chooses a loss function such as cross-entropy, importance sampling or CISPO.[12] A saved checkpoint can be turned into a dedicated deployment that answers requests through an endpoint compatible with the OpenAI API.[12]

River publishes two worked examples on its homepage. One is a minimal GRPO loop that trains Kimi K2.6 on GSM8K math word problems. The other follows the ScaleRL recipe from the 2025 paper "The Art of Scaling Reinforcement Learning Compute for LLMs" by Devvrit Khatri and colleagues, training Qwen3.6 35B on the Polaris-53K math dataset; River says the run cost under $1,000.[12][22] The company says it stress-tests the API to keep KL mismatch and numerical instability under control, problems that can undermine a reinforcement learning loop at scale.[12]

TechCrunch wrote that the API was intended to be "an antidote to prompt engineering," and that River was promising, through its neocloud offering, to supply the post-training expertise enterprises need as they seek more control over the models they use.[3]

River Cloud and River Cluster

River sells the same software in two deployment forms. River Cloud is managed by River and billed per token for training, sampling and inference, with checkpoint storage charged at $0.10 per gigabyte per month; customers can download their trained weights. River Cluster deploys the same API, workflows and Console on GPUs a customer already owns, keeping data and weights inside the customer's environment, and is priced on request.[12][15] The River Console shows a team's training runs, checkpoints, deployments, usage and API keys.[12] River's security page says its own services run on Google Cloud Platform.[16]

River Inference

River Inference, launched on October 11, 2026, lets customers call hosted open-weight models through an OpenAI-compatible API using their River API key.[14] It started with DeepSeek V4.1-Flash and GLM-5.3-Flash. River's changelog said the service was 20-60% cheaper than other providers, and Babuschkin wrote that the two models were 20% cheaper than the offerings from DeepSeek and Z.ai. He said River had written its own training and inference stack from scratch in Rust, which made it much more efficient, and offered new sign-ups $25 in free credits.[14][18] River's homepage compares its prices with other providers:[12]

Model	Provider	Input (per 1M tokens)	Output (per 1M tokens)
DeepSeek V4.1 Flash	River	$0.12	$0.48
DeepSeek V4.1 Flash	DeepSeek	$0.15	$0.60
DeepSeek V4.1 Flash	Fireworks AI	$0.30	$1.20
DeepSeek V4.1 Flash	Together AI	$0.30	$1.20
GLM-5.3 Flash	River	$0.12	$0.40
GLM-5.3 Flash	Z.ai	$0.15	$0.50
GLM-5.3 Flash	Fireworks AI	$0.15	$0.50
GLM-5.3 Flash	Together AI	$0.15	$0.50

Prices are as listed by River in October 2026 and are described by the company as preview rates subject to change.[12][15]

Supported models

As of October 2026, River's website listed the following open-weight models as optimized for both training and inference, each with a 262,000-token context option:[12]

Family	Model	Architecture
Qwen	Qwen3.8 27B	Dense
Qwen	Qwen3.6 35B	Mixture of experts
Qwen	Qwen3.5 397B	Mixture of experts
Kimi	Kimi K2.6	Mixture of experts
GLM	GLM 5.2	Mixture of experts
GLM	GLM-5.3-Flash	Mixture of experts
DeepSeek	DeepSeek V4.1-Flash	Mixture of experts
Nemotron	Nemotron 3.5 Super VL	Mixture of experts

River Recipes and the text-to-SQL study

On October 6, 2026, River introduced River Recipes, an open-source project that turns published research into tested, reproducible training recipes. The recipes are hosted on GitHub under the Apache 2.0 license.[19][30] The first recipe came from a study of text-to-SQL, the task of turning a natural-language question into a database query.[20]

The study started from ReViSQL, a method described in a 2026 arXiv paper by Yuxuan Zhu, Tengjun Jin, Yoojin Choi and Daniel Kang. The paper argued that human-level text-to-SQL is reachable by fine-tuning an LLM with reinforcement learning with verifiable rewards (RLVR) on cleaned data. Its authors built BIRD-Platinum, a verified subset of the BIRD training set, and reported that a Kimi K2.6 model trained with their reward shaping reached 92.96% on Arcwise-Plat, an expert-verified version of BIRD.[21]

River reported that the authors' released code, originally written for the Tinker training API from Thinking Machines Lab, collapsed in longer runs around the second or third epoch. It ran the unmodified code on River, Tinker and Fireworks, found that every run fell short of the paper's result after one epoch, and then rebuilt the training loop. Its changes included asynchronous reinforcement learning that overlaps sampling and training, a stricter grader, an updated version of the VeriEQL equivalence checker, and a fix for a parsing bug: when Kimi K2.6 hit the token limit mid-reasoning, the original code treated the unfinished reasoning as the answer and could grade it as correct. River said it suspected that bug was one of the main reasons the reported Kimi results could not be reproduced with the original code.[20]

River also built BIRD-Diamond, a further cleaned version of BIRD-Platinum, and reported that training on it lowered evaluation scores by about 1% on average. Its analysis was that the extra cleaning removed evidence errors that had taught models some skepticism, leaving them too trusting of the evidence supplied with the test questions.[20] With its recipe, River said, four open models reached the 92.96% accuracy that ReViSQL had described as human-level, and five trained open models outscored the strongest proprietary model it measured, GPT-6 Astra Pro at maximum effort, by an average of 4.0 points at less than 2% of the cost per question. GLM-5.3 Flash scored highest, 4.6 points above GPT-6 Astra Pro at 0.5% of its cost. River also tested the models on the 135 SQLite tasks of Spider2-Lite, part of the Spider 2.0 benchmark: every model improved after training, but River's best models finished just short of the proprietary frontier.[20] These results are the company's own and used grading rules matching the ReViSQL authors'.[20]

On October 15, 2026, River said it had worked with Nvidia to support Nemotron 3.5 Super VL on the API from its release day. Using the text-to-SQL recipe, River reported that the model scored 80.7% on Arcwise-Plat before training and 92% after, ahead of every untrained proprietary model it evaluated.[23]

Hardware and personal AI plans

River's vision page lays out four layers: interfaces (in development), models and algorithms (research), infrastructure (the River API, available now) and personal hardware (research and development). It says the company is working toward "powerful personal hardware to run AI locally," with hardware and software designed together.[13] The company's August announcement likewise described "new hardware that lets personal AI live close to you, running for you rather than in someone else's data center."[2]

River's job listings give more detail. As of October 2026 its Greenhouse board listed hardware roles in Palo Alto and Austin for RTL design, physical design through tape-out, design verification, kernel engineering and compilers, describing a custom AI accelerator and system-on-chip components on "advanced foundry nodes." The compiler role calls for lowering PyTorch models to a custom instruction set using MLIR and LLVM.[25] SiliconANGLE reported the same plan from a job posting, describing a custom system-on-chip with an onboard machine learning accelerator.[26] River's research listing asks for work on continual learning, so that models adapt to local user interaction without catastrophic forgetting, and on foundation models that run efficiently on the company's own personal hardware.[25]

Mission and approach

River's stated mission is "to create personal AI owned and shaped by each individual." Its job listings say the company is "rewriting the entire stack from scratch: personal hardware for local inference, custom training infrastructure, next-generation UIs, and frontier deep learning research."[25] In its funding announcement, River argued that most companies run general-purpose models trained for the broadest possible audience, and that building a custom model has required a dedicated infrastructure team, specialized hardware and months of work.[2]

The company presents ownership as covering three things: the data an AI learns from, the model weights, and the hardware it runs on.[13][31] Babuschkin has tied this to AI alignment. In the launch post he wrote that "an AI that knows you deeply, understands your goals, and can be trusted to act in your interest is, by definition, highly aligned," and that the company does not treat alignment and safety as separate problems.[1] The funding announcement put the same idea as aligning AI "directly to each person" rather than aligning one model to billions of users.[2] He has described the agents he expects as "less like the assistants you call on today when you need a task done, more like guardian angels."[1]

At the funding announcement Babuschkin said: "AI should be open, freely available, and affordable. It should feel like it is working for the person using it, not the lab that trained it."[2] General Catalyst CEO Hemant Taneja framed the investment in national terms: "American leadership in AI urgently requires leadership in open-weight models, while maintaining a lead in closed frontier models." He called River's agenda "a priority for American resilience."[2][5]

Customers

River's homepage lists Perplexity, Ivo, Abacus AI, Mixedbread and Afori as customers, under the heading "Trusted by teams at the frontier."[12] The company has not published the terms of these relationships, and the list comes from River itself. In August 2026, AI Weekly noted that neither the funding announcement nor the press coverage had named a first paying customer.[28]

Reception

TechCrunch called the round "an eye-popping-size investment for a nascent company, and perhaps another indication of the overheated AI atmosphere," while adding that River's premise arrived as enterprises wanted more control over their AI models, including open-weight ones. It concluded that how River's technology would differ "remains to be seen."[3] The Next Web pointed out that River's speed and cost claims had not been independently tested.[5]

A June 2026 analysis by the Chinese outlet Xinzhiyuan, republished by 36Kr, called River's move to align AI with individual users both its main advantage and the first question it would face, asking who would take responsibility for group safety and wider social risks if each AI were aligned only to its owner.[9]

Debarghya Das's "The Ultimate Neolab List," a ranking of AI research labs by valuation, includes River, classifying it as a frontier lab working on personalized AI.[29]

See also

References

Babuschkin, Igor. "Introducing River AI." River AI blog, June 10, 2026.
River AI. "River AI raises $1.1B in funding across Series Seed and Series A." River AI blog, August 11, 2026.
Bort, Julie. "General Catalyst leads $1.1B round into 2-month-old River AI." TechCrunch, August 11, 2026.
Jayaprakash, Prathik. "XAI co-founder's startup River AI raises $1.1 billion to expand custom AI tools." Reuters, via Yahoo Finance, August 11, 2026.
Stanciuc, Ana-Maria. "River AI raised $1.1bn to let companies train and keep their own models." The Next Web, August 11, 2026.
Shrivastava, Rashi, and Anna Tong. "Un ex socio de Elon Musk busca US$ 1.000 millones para crear su 'neolaboratorio' de IA." Forbes Argentina, May 15, 2026 (Spanish translation of a Forbes.com report first published May 14, 2026).
Babuschkin, Igor (@ibab). "We're launching River AI, a new AI company with the mission to build AI systems that are owned and shaped by you." X, June 10, 2026.
Zeff, Maxwell. "Co-founder of Elon Musk's xAI departs the company." TechCrunch, August 13, 2025.
Xinzhiyuan. "马斯克前联创洽谈10亿美元巨融，要造一个真正归你的AI" (Musk's former co-founder in talks for a $1 billion raise to build an AI that truly belongs to you). 36Kr, June 22, 2026.
Razin, Aleksei. "Успевший поработать в xAI, OpenAI и Google Игорь Бабушкин основал собственный ИИ-стартап" (Igor Babuschkin, who worked at xAI, OpenAI and Google, founded his own AI startup). 3DNews, June 11, 2026.
River AI. "API: v0.1 preview." River AI website, archived by the Internet Archive on June 10, 2026.
River AI. "River AI: Your own frontier AI lab. In one API." River AI homepage, accessed October 27, 2026.
River AI. "Our vision." River AI website, accessed October 27, 2026.
River AI. "Changelog." River AI website, accessed October 27, 2026.
River AI. "Pricing." River AI website, accessed October 27, 2026.
River AI. "Security at River." River AI website, accessed October 27, 2026.
River AI. "River handbook: Train LLMs and agents with River." River API documentation, accessed October 27, 2026.
Babuschkin, Igor (@ibab). "We've launched fast production inference for open-weight models!" X, October 11, 2026.
River AI. "Introducing River Recipes." River AI blog, October 6, 2026.
River AI. "Achieving Frontier Text-to-SQL with the River API." River AI blog, October 6, 2026.
Zhu, Yuxuan, Tengjun Jin, Yoojin Choi, and Daniel Kang. "Human-Level Text-to-SQL via Reinforcement Learning on Verified Data, Without Pipeline Engineering." arXiv:2603.20004, March 2026.
Khatri, Devvrit, et al. "The Art of Scaling Reinforcement Learning Compute for LLMs." arXiv:2510.13786, October 2025.
River AI. "Announcing Nemotron 3.5 Super VL on the River API." River AI blog, October 15, 2026.
River AI. "Careers." River AI website, accessed October 27, 2026.
River AI. "River AI job board." Greenhouse, accessed October 27, 2026 (listings including "RTL Design Engineer, Hardware," "Compiler Engineer, Hardware," "Physical Design Engineer, Hardware" and "Research Engineer / Research Scientist").
Deutscher, Maria. "Personalized AI startup River AI raises $1.1B from consortium backed by Nvidia, AMD." SiliconANGLE, August 11, 2026.
Dealroom. "Igor Babuschkin's River AI raises $1.1B to build an open AI stack." Dealroom News, August 11, 2026.
Dufresne, Alexis. "Babuschkin's River AI closes $1.1B at ~$5B for user-owned AI." AI Weekly, August 11, 2026.
Das, Debarghya. "The Ultimate Neolab List." debarghyadas.com, accessed October 27, 2026.
River AI. "river-recipes." GitHub repository, accessed October 27, 2026.
River AI (@river_ai_inc). "Today, we're sharing that River AI has raised $1.1 billion." X, August 11, 2026.

Editing: River AI