AI Wiki

Explore AI Wiki

Citation and evidence

Imbue

27 min full readUpdated 55 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI AgentsAI Companies

Cite this article

Imbue is a San Francisco artificial intelligence company co-founded in 2021 by Kanjun Qiu (chief executive officer) and Josh Albrecht (chief technology officer). It launched as a research lab named Generally Intelligent, which studied reinforcement learning agents and the theory of deep learning, and renamed itself Imbue in September 2023 when it announced a $200 million Series B round at a valuation of more than $1 billion.[1][2][3] In 2024 the company described pre-training a 70B-parameter language model on its own 4,088 GPU NVIDIA H100 cluster and open-sourced the infrastructure scripts, evaluation datasets and the Bayesian Optimization based hyperparameter optimizer CARBS that came out of that work.[5][6][19] Since 2025 its public output has been mostly open-source software: Sculptor, a desktop app for running coding agents in parallel, followed in 2026 by command-line developer tools (mngr, Vet, Blueprint, Latchkey, Keystone, Offload), the X feed filter Bouncer, the self-hosting service Cloud in a Bottle, and Imbue Studio, a "personal AI operating system" announced on 1 October 2026 and offered through a waitlist.[7][8][42] The company says its tools "should work for you, not the company that built them."[8] Imbue had not announced any funding round after its October 2023 Series B extension as of October 2026.[28][54]

Related: Reasoning (artificial intelligence)NvidiaClaude CodeAnthropic

Imbue logo
Imbue logo, from the official website imbue.com. Trademark of its owner.

Company overview

ItemDetail
Legal nameImbue, Inc.[55]
Prior nameGenerally Intelligent (2021 to September 2023)
Co-foundersKanjun Qiu (CEO), Josh Albrecht (CTO)
HeadquartersSan Francisco, California
Founded2021[23]
RenamedSeptember 2023
Initial funding (2022)More than $20 million contributed plus more than $100 million committed in options and milestone-based drawdowns[26]
Series B (initial)$200 million, announced September 7, 2023
Series B (extension)Additional $12 million from the Amazon Alexa Fund and Eric Schmidt, announced October 2023
Reported valuationOver $1 billion (Series B, September 2023); no later valuation disclosed as of October 2026
Series B leadAstera Institute[14]
Named Series B participantsNvidia, Kyle Vogt, Simon Last (initial round); Amazon Alexa Fund, Eric Schmidt (extension)[1][28]
Other named backersTom Brown, Drew Houston, Jonas Schneider, Arash Ferdowsi, Jed McCaleb[10][26][28]
Compute partnersVoltage Park (cluster management), Dell (PowerEdge servers), Nvidia (4,088 H100 GPU cluster used for the 70B run)[4][5]
Products (October 2026)Imbue Studio (waitlist), Sculptor, mngr, Vet, Blueprint, Bouncer, Cloud in a Bottle[8]
Open-source releasesAvalon, CARBS, cluster-health, evaluation datasets, Sculptor, mngr, Vet, Latchkey, Keystone, Darwinian Evolver, Imbue Catalyst
PodcastGenerally Intelligent (since December 2020)[11]

Expanded article table

History

Founders and earlier companies

Qiu and Albrecht co-founded Sourceress, an AI recruiting company that went through Y Combinator in the Summer 2017 batch.[9] Qiu says Sourceress raised $13 million; she shut it down before starting Generally Intelligent and told TechCrunch the decision reflected the high-churn nature of the leads-sourcing business.[24][10] Before that, Qiu was chief of staff at Dropbox, a period in which she says the company grew from 300 to 1,500 people, and she co-founded Ember Hardware, which designed laser displays for virtual reality headsets.[10][24] TechCrunch also reported that she had worked as an algorithmic trader for two years; her own site says she co-authored Sew Electric, a book that teaches children programming with sewable electronics, and tinkered with hardware at the MIT Media Lab.[10][24] Albrecht's earlier companies included BitBlinder, a privacy-preserving torrenting tool, and CloudFab, a 3D printing services company.[10] Qiu and Albrecht are also general partners, with managing partner Ali Rohde, at Outset Capital, an early-stage fund that writes small checks into startups.[25][24]

Origins as Generally Intelligent

Qiu and Albrecht launched the Generally Intelligent podcast on December 6, 2020, describing it as a podcast "made by deep learning researchers, for deep learning researchers."[11] Guests have included Percy Liang, Chelsea Finn, Noam Brown, Jonathan Frankle, Tri Dao, Jamie Simon and Rylan Schaeffer, and a 38th episode on the theory of deep learning was published in April 2026.[54][48]

The company itself was founded in 2021.[10][23] In a 2023 interview Qiu said the founders had been thinking seriously about it since late 2019 or early 2020, when they saw that scaling language models was starting to work.[18] Generally Intelligent emerged from stealth on October 20, 2022. Its launch post said backers had contributed more than $20 million and committed more than $100 million in future funding "in a combination of options and technical milestones", and named Jed McCaleb (founder of the Astera Institute), Tom Brown (lead author of GPT-3), Jonas Schneider (former robotics lead at OpenAI) and Drew Houston as investors.[26] TechCrunch added Dropbox co-founder Arash Ferdowsi and the Astera Institute to the investor list, reported a team of about 12 people, and named Neuralink founding team member Tim Hanson as a board member.[10] Qiu told TechCrunch that the $100 million agreement was "guaranteed through a drawdown setup" with triggers the company would not disclose.[10]

The lab described itself as "primarily a research organization" that did "not include medium-term commercialization milestones" and aimed to develop "generally capable AI agents with human-like intelligence."[26] Its first major release was Avalon, a procedurally generated 3D reinforcement learning benchmark accepted to the NeurIPS 2022 Datasets and Benchmarks track.[12][13] In Avalon, embodied agents must survive by navigating terrain, hunting or gathering food and avoiding hazards; the reward function, world dynamics and action space are the same for all 20 tasks, which range from "eat" and "throw" to "hunt" and "navigate".[12] The paper lists Albrecht as first author and Qiu as last author. The launch post framed the tasks as the kinds of problems that "in an evolutionary sense, led to the emergence of our species' unique cognitive abilities" and noted that Avalon is built on the open-source Godot game engine.[26] Qiu later said the team concluded that pure reinforcement learning was "not a good vehicle for planning and reasoning", which pushed it toward language-based reasoning.[18]

Rebrand and Series B

On September 7, 2023, Generally Intelligent announced a $200 million Series B "at a valuation of over $1 billion" and renamed itself Imbue, a name meant to reflect a focus on "imbuing computers with intelligence and human values."[27] Participants named in the announcement were the Astera Institute, Nvidia, Cruise CEO Kyle Vogt and Notion co-founder Simon Last; Reuters and Crunchbase News reported the same round and valuation, and TechCrunch reported that it brought Imbue's total funding to $220 million.[1][2][3][27] Maginative later described the Astera Institute as the round's lead.[14] Crunchbase News placed the deal among Nvidia's 2023 investments in AI startups, alongside rounds for AI21 Labs and Adept AI.[2]

The announcement post laid out a "full stack" research agenda with five parts: pre-training its own "very large (>100B parameter) models" optimized for internal reasoning benchmarks on a roughly 10,000 GPU H100 cluster; prototyping agents, mainly agents that write code, for internal use; inventing new interfaces; building tools such as CARBS; and developing a theory of deep learning.[27] TechCrunch summarized the shift as a move away from training agents in 3D worlds toward models the company found "internally useful", including models that can code, and noted that Imbue did not intend to productionize much of that work at the time.[1]

In October 2023 Imbue added $12 million from the Amazon Alexa Fund and former Google chief executive Eric Schmidt, which the company said extended the Series B "to over $210m".[28][14] Imbue's announcement quoted Paul Bernard, director of the Alexa Fund, and listed Kyle Vogt, Drew Houston, Tom Brown and Simon Last as other backers.[28] Maginative, reporting the extension, wrote that commercial applications of Imbue's technology were "still possibly years away" and that it remained to be seen whether its internal demos justified the valuation.[14]

Compute build-out and the 70B project (2023 to 2024)

On November 28, 2023, Dell and Imbue announced a $150 million agreement for a cluster of Dell PowerEdge XE9680 servers with "nearly 10,000" NVIDIA H100 GPUs, designed with smaller clusters for fast experimentation as well as fast networking into one large cluster. The release said the system was managed by Voltage Park and that Imbue was already using it to train models and build prototype agents that fix bugs in code and analyze long documents.[4] Imbue's later technical write-up described one training cluster of 4,088 H100 GPUs on 511 hosts and thanked Voltage Park, Dell, H5 and NVIDIA for help setting it up.[5]

On June 25, 2024, Imbue published a series of posts saying that "earlier this year" it had pre-trained a 70B-parameter model on 2 trillion tokens and fine-tuned it on multiple-choice reasoning benchmarks.[6] Albrecht discussed the work the same day on the Latent Space podcast, as a joint guest with Databricks chief AI scientist Jonathan Frankle.[15][16]

Products and reorganization (2025 to 2026)

In April 2025 Imbue released Sculptor as an early research preview: a sandboxed coding agent environment that checked code for issues such as missing tests, hardcoded variables and race conditions, let users launch agents to fix them in parallel, and supported custom checks written as LLM prompts or shell commands.[29] After users asked for ways to run several Claude Code agents safely in parallel, the company rebuilt Sculptor around that use case and announced the new version in a blog post dated September 26, 2025.[7] Albrecht introduced it on Hacker News on September 30, 2025 as "a desktop app that lets you safely run Claude Code agents by putting them in separate docker containers."[30] Imbue's later posts date the launch to early October 2025.[31][32]

In a January 2026 post titled "A more radical Imbue", Albrecht wrote that he and Qiu had become "disillusioned" during the 2025 sprint to ship Sculptor, which he said had attracted users who kept using it week after week, and that the company had fallen into "default startup patterns" such as feature roadmaps and KPIs.[31] He announced a shift toward "more, smaller open software projects, rather than focusing on one top-down product", with individual employees proposing and owning projects tied to the mission of "making tech serve humans."[31] Through 2026 Imbue released a series of open-source developer tools (mngr, Vet, Latchkey, Keystone, Offload, Blueprint), consumer tools (Bouncer, Cloud in a Bottle), and research tools (Darwinian Evolver, Imbue Catalyst), described in the sections below.[8]

On October 1, 2026, Imbue announced Imbue Studio and published an essay by Albrecht, "It's time for Personal Computing 2.0", arguing that users should own their data and be able to move between AI providers.[42][43] Studio was available by waitlist at launch.[42]

Policy work

Imbue runs a policy program led by Matt Boulos, its head of policy and safety; its home page says the work aims to build "legal frameworks that protect individual liberty" as AI companies accumulate power.[8][51] In late October 2025 Imbue was one of the founding members of the Agentic Futures Initiative, an industry group on AI agent policy organized by the lobbying firm Aquia Group, alongside Anthropic, Intuit, Edgerunner AI, Unstructured, Shift5 and Lilt AI.[50][51] Semafor reported that Boulos hoped agents would develop common protocols "much like the early web."[51] Qiu discussed open versus closed AI ecosystems on The Verge's Decoder podcast in October 2025.[52]

Mission and research thesis

Imbue's about page argues that the software and agents people use today "work for the company that built them" rather than for their users, and says the company is building "an open agent ecosystem where you can understand, modify, and verify your agents."[23] The about page says the company was founded "to build an AGI future where humans remain at the helm."[23]

The research thesis behind the 2023 rebrand was that reasoning is "the primary blocker to effective AI agents."[27] The company said it would create pre-training data "specifically designed to reinforce good reasoning patterns" and develop techniques that spend far more compute at inference time.[27] In a 2023 Latent Space interview, Qiu said the company generates "very specific things that are like reasoning traces" as training data and described code as "the most explicit example of reasoning data on the internet."[18] The 2023 announcement listed five threads of work:[27]

  1. Models: pre-training large models optimized for internal reasoning benchmarks.
  2. Agents: prototype agents used internally, mainly agents that write code.
  3. Interfaces: new ways for people to work with agents, on the view that "today's AI chat interfaces are skeuomorphic".
  4. Tools: internal tooling such as debugging interfaces and the CARBS hyperparameter optimizer.
  5. Theory: a theory of deep learning, with research then focused on feature learning.

The theory thread has continued. Imbue says it has for several years funded a small research team called Feature Lab, led by research fellow Jamie Simon.[48] In April 2026 Simon, Daniel Kunin and 12 co-authors posted "There Will Be a Scientific Theory of Deep Learning" to arXiv, a paper that proposes the name "learning mechanics" for an emerging theory of how neural networks train.[48][49]

Infrastructure and the 70B project

The main public record of Imbue's model training is the June 2024 series: an overview post and three detailed posts on evaluations, infrastructure and hyperparameter optimization.[5][6][19] Imbue said the work was done by "about a dozen" of its engineers and researchers.[6]

Cluster topology

The cluster described in the infrastructure post had 4,088 H100 GPUs spread across 511 computers, eight GPUs per computer; the host count was 511 because some connections were reserved for the Unified Fabric Manager (UFM) nodes that managed the InfiniBand network.[5] Each GPU was connected directly to a ConnectX-7 card that could transmit and receive at 400 Gbps simultaneously, and the three-tier InfiniBand network was "fully non-blocking", with 320 network switches in the fabric and more than 12,000 connections overall.[5] Training traffic ran over InfiniBand; a separate 100 Gbps Ethernet network carried datasets and checkpoints, and a management network gave access to each machine's baseboard management controller (BMC) for remote monitoring, BIOS settings and power management.[5]

Bring-up and cluster-health tooling

The infrastructure post described bringing the cluster up from bare metal: about 10% of machines initially failed to boot, mostly because of physical problems, internal PCIe cables had to be reseated on around a quarter of hosts, and a miswired top level of the fabric had left eight disconnected networks until it was rewired.[5] Imbue released the scripts it developed in the imbue-ai/cluster-health repository under the MIT License.[5][20] The release includes:[5][20]

  • Host-level health checks that confirm a host is free of known errors.
  • A GPU stress test that confirms GPUs can allocate large tensors and perform standard operations.
  • Networking tests for GPU-to-GPU communication within a machine (NVLink) and across machines (InfiniBand).
  • A script that parses the UFM event log and decides which network ports to disable.
  • An InfiniBand burn-in workload (ib_burn) that aims to exercise every available link.
  • A patch to NCCL that improves logging around errors and stalls.

The post gave a rule of thumb that about 3% of machines in a new GPU cluster break every week, adding that the failures cluster in a small number of problem machines rather than being uniform.[5] It also traced periodic throughput dips, which affected a growing share of operations as training progressed, to Python's automatic garbage collection running at different times on different workers of the synchronous FSDP job, which the team fixed by scheduling garbage collection at fixed intervals across all hosts.[5] Albrecht and Frankle compared notes on both issues in the Latent Space episode.[15]

CARBS hyperparameter optimizer

CARBS (Cost-Aware Pareto-Region Bayesian Search) was first published in June 2023 and is available at imbue-ai/carbs under the MIT License.[21][22][53] It models both performance and cost (in GPU-hours) as Gaussian Processes and searches locally around the performance-cost Pareto frontier rather than over a fixed global search space, which lets it learn how each hyperparameter should scale with compute.[22]

Reported results include:[22]

  • Reproducing the Chinchilla scaling laws result on model size versus training tokens from 340 training runs of models between 19 million and 1.5 billion parameters, while tuning 19 hyperparameters at once. The original DeepMind analysis trained 50 models of 44 million to 16 billion parameters.
  • On OpenAI's ProcGen suite, a CARBS-tuned PPO baseline matched the previous best performance at a quarter of the cost and scored more than 16% higher than the previous state of the art.

For the 70B project, Imbue used CARBS to tune dozens of hyperparameters over thousands of small-scale experiments, both to choose settings for the large run and to predict its performance.[19]

The 70B model and evaluations

Imbue's 70B model was a dense transformer with grouped-query attention, SwiGLU activations, RMS normalization and a custom tokenizer, pre-trained on 2 trillion tokens.[6] The company said that using CARBS it scaled to 70B parameters "on our first attempt with minimal training instability and no loss spikes."[6] After fine-tuning on multiple-choice reasoning benchmarks, Imbue reported that the model outperformed GPT-4o evaluated zero-shot and approached a fine-tuned Llama 3 70B, which had been pre-trained on more than seven times as much data.[6]

Alongside the results, Imbue released:[6]

  • Cleaned-up and extended versions of 11 popular NLP reasoning benchmarks.
  • CodeComprehension, a new code-focused reasoning benchmark.
  • A fine-tuned 70B model, built with Meta Llama 3, for measuring question quality.
  • A dataset of 450,000 human judgments about question quality.
  • The infrastructure scripts and CARBS.

The weights of the 70B base model were not part of the released toolkit.[6]

Products

Sculptor

Sculptor is an open-source desktop app for running coding agents in parallel, released under the MIT License and available for Apple Silicon Macs and Linux.[17][33] The September 2025 version ran each agent in its own container, so agents could execute code without touching the developer's machine, and introduced:[7]

  • Pairing Mode, which syncs an agent's work from its container into the local repository in both directions, so the developer can test and edit it in their own IDE.
  • Session history, which saves each agent's plans, chats, tool calls and code changes so a session can be reopened later.
  • Merge conflict detection, which flags potential conflicts between agents and can hand them back to an agent to resolve.
  • Suggestions (beta), which points out issues in an agent's code.

Imbue said in December 2025 that it had added support for OpenAI's Codex alongside Claude Code.[7] As of October 2026, the product page describes each workspace as an isolated git worktree with its own branch, terminal and diff view, says Sculptor runs locally and requires no subscription beyond the user's model provider, and lists built-in workflow skills for spec writing, mock generation and test-driven bug fixing.[33] The GitHub README says Sculptor ships with integrated support for the Pi agent harness in addition to Claude Code and can manage other terminal-based agents, while still describing the project as "an experimental research preview."[17]

Developer tools (2026)

ToolAnnouncedDescription
LatchkeyFebruary 2026Command-line tool that lets agents call HTTP APIs with ordinary curl commands and injects stored credentials, presented as a simpler alternative to the Model Context Protocol; MIT-licensed[36]
VetMarch 2026Local code review tool that reads an agent's conversation history alongside its code changes to catch cases where the agent's actions do not match the request; runs from the CLI, in CI or as an agent skill[35]
OffloadMarch 2026Rust CLI that spreads a test suite across 50 to 200 or more cloud sandboxes on Modal; Imbue reported more than a 6x speedup on its own integration tests[32]
KeystoneMarch 2026Python CLI that has a coding agent write a working Dockerfile and dev container for an arbitrary repository inside a Modal sandbox[38]
mngrApril 2026Unix-style CLI for launching and managing many coding agents (Claude, Codex or others) on a laptop, in Docker, on Modal or on any SSH host; MIT-licensed[34]
BlueprintApril 2026Planning tool, available as agent skills and as an extension for VS Code, Cursor and Windsurf, that reads a codebase and asks the developer multiple-choice questions before writing a plan[37]

Expanded article table

Bouncer and Cloud in a Bottle

Bouncer, announced in April 2026 and available for iPhone, Chrome and Firefox, filters an X (Twitter) feed using plain-language filter phrases. Imbue's announcement says it classifies posts with the open model Gemma 4 26B-A4B served from its own data center.[39] Bouncer 2.0, in July 2026, added detection of AI-generated text and images using purpose-trained models; Imbue reported 99.8% accuracy at separating human from AI-generated text on its own test set with a Qwen 3 4B backbone, and said its iOS app runs text detection on the device using a LoRA adapter on Gemma 4 E2B.[40]

Cloud in a Bottle is a service for self-hosting personal and open-source apps on the user's own hardware, a rented server or an Imbue-managed instance, with a catalog that includes Nextcloud, Jellyfin and VaultWarden.[41]

Imbue Studio

Imbue Studio, announced on October 1, 2026, lets users build "studios" in which they work with AI agents and other people to make personal tools from services such as email, calendar and Slack, and share them by link.[42] Imbue says Studio is open source and can modify its own code, lets users switch between Anthropic, OpenAI or open-model subscriptions mid-conversation, runs in Imbue's cloud or on the user's own machine, encrypts traffic and stored data end to end, and runs agents in a sandbox with zero access by default.[42] At announcement it was available by waitlist only.[42]

Research since 2025

  • Code evolution and ARC-AGI-2. In February 2026 Imbue open-sourced the Darwinian Evolver, an LLM-driven code evolution tool inspired by Sakana AI's Darwin Gödel Machines, which it first built to optimize Vet.[45] Applied to ARC-AGI-2, the method raised Kimi K2.5 from 12.1% to 34.0% at $2.67 per task, which Imbue called a record for open-weight models, and reached 95% with Gemini 3.1 Pro at $8.71 per task. These scores are on the public evaluation set; Imbue said its semi-private submission had not been evaluated because new community submissions were on hold.[44]
  • Imbue Catalyst. In July 2026 Imbue released Imbue Catalyst under the AGPL-3.0 license, an evolution-based tool for computational research.[46] On Andrej Karpathy's AutoResearch nanochat setup, Imbue reported a validation score of 0.9361 bits per byte after 340 experiments and compared the run with Recursive Superintelligence's reported results, noting that the number of experiments in Recursive's run was estimated.[46] A second post used Catalyst to search for mechanistic explanations of a neuron-clustering effect seen when training shallow networks.[47]
  • Theory of deep learning. See the Feature Lab paper described above.[48][49]

Significance

Dell's 2023 announcement called Imbue "one of the few independent AI labs that develops its own foundation models."[4] Its June 2024 posts described a from-scratch 70B pre-training run in unusual detail, down to hardware failure rates, wiring faults and the remediation scripts it released, and Latent Space devoted an episode to the release.[5][15] Unlike frontier labs such as OpenAI, Anthropic and Google DeepMind, it did not release a general-purpose chat model, and by 2026 its public work had moved to open-source agent tooling and personal software built on other companies' models.[6][8] The Avalon launch post called it "the first reinforcement learning benchmark of its kind" for learning in complex procedurally generated 3D worlds.[26]

Comparison with peer companies

For comparison, other agent- and code-focused AI startups of the early 2020s include:

CompanyFocus
ImbueReasoning-focused models (to 2024), then open-source agent tools and personal software
Adept AIAction transformers for web and desktop UIs; later partial team move to Amazon
Inflection AIConsumer personal AI (Pi assistant); pivoted to enterprise after 2024 Microsoft deal
Magic.devLong-context code generation models
Cognition AIDevin software engineer agent and Windsurf-derived IDE

Expanded article table

Adept's Series B was a $350 million round in March 2023 at a post-money valuation of at least $1 billion, six months before Imbue's.[2]

Limitations and criticisms

  • Unreleased weights and narrow comparisons. Imbue did not release the 70B base model's weights, and its headline comparison with GPT-4o covers multiple-choice reasoning benchmarks on which Imbue's model was fine-tuned while GPT-4o was evaluated zero-shot and without chain-of-thought. Imbue itself noted that this "does not reflect the best possible scores" GPT-4o can achieve.[6]
  • Commercial uncertainty. When Imbue extended its Series B in 2023, Maginative questioned whether internal demos justified a billion-dollar valuation for a company whose commercial applications were "still possibly years away."[14] Albrecht's 2026 post described internal worries during 2025 about whether a research company could ship a product with users.[31]
  • Self-reported results. The ARC-AGI-2 scores Imbue reported in 2026 are on the public evaluation set and had not been verified on the semi-private set used by the official leaderboard.[44]
  • Coding agents. Sculptor and mngr sit in the same design space as Devin from Cognition AI, Claude Code from Anthropic, OpenAI Codex and other AI coding agents.
  • Reinforcement-learning environments. Avalon belongs to the tradition of procedurally generated benchmarks alongside OpenAI Gym derivatives and SIMA from DeepMind.
  • Open-source training infrastructure. The cluster-health scripts overlap with vendor tools such as NCCL tests and CUDA diagnostics, focused on H100 and InfiniBand clusters.
  • Hyperparameter optimization and scaling. CARBS applies hyperparameter tuning and Bayesian Optimization to a Pareto-frontier setting tied to scaling laws.
  • Evolutionary search with LLMs. The Darwinian Evolver and Catalyst build on Sakana AI's Darwin Gödel Machine and Karpathy's AutoResearch.[45][46]

Continue reading

More from See also

Topics:AI AgentsAI Companies

References

  1. ^1 ^2 ^3 ^4Kyle Wiggers, "Imbue raises $200M to build AI models that can 'robustly reason'", TechCrunch, 2023-09-07. techcrunch.com/...-models-that-can-robustly-reason. Accessed 2026-10-27.
  2. ^1 ^2 ^3 ^4Chris Metinko, "AI Lab Imbue Gets $200M From Nvidia, Others; Hits $1B Valuation", Crunchbase News, 2023-09-07. news.crunchbase.com/...unicorn-imbue-astera-nvidia. Accessed 2026-10-27.
  3. ^1 ^2Reuters, "Imbue raises $200 million at over $1 billion valuation in AI funding boost" (syndicated by Investing.com), 2023-09-07. investing.com/...ation-in-ai-funding-boost-3170172. Accessed 2026-10-27.
  4. ^1 ^2 ^3Dell Technologies, "Imbue to Develop Next-Generation AI Models with $150 Million Dell High Performance Computing System", PR Newswire, 2023-11-28. prnewswire.com/...mance-computing-system-301998808. Accessed 2026-10-27.
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Imbue, "From bare metal to a 70B model: infrastructure set-up and scripts", Imbue Research, 2024-06-25. imbue.com/...70b-infrastructure. Accessed 2026-10-27.
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Imbue, "Training a 70B model from scratch: open-source tools, evaluation datasets, and learnings", Imbue Research, 2024-06-25. imbue.com/...70b-intro. Accessed 2026-10-27.
  7. ^1 ^2 ^3 ^4Kanjun Qiu, "Sculptor: the missing UI for parallel coding agents", Imbue Blog, 2025-09-26 (updated with a December 2025 note on Codex support). imbue.com/...sculptor-announce. Accessed 2026-10-27.
  8. ^1 ^2 ^3 ^4 ^5 ^6Imbue, "We build honest software (company home page)", Imbue. imbue.com. Accessed 2026-10-27.
  9. ^Y Combinator, "Sourceress", Y Combinator Companies Directory. ycombinator.com/...generally-intelligent. Accessed 2026-10-27.
  10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Kyle Wiggers, "Generally Intelligent secures cash from OpenAI vets to build capable AI systems", TechCrunch, 2022-10-20. techcrunch.com/...vets-to-build-capable-ai-systems. Accessed 2026-10-27.
  11. ^1 ^2Josh Albrecht and Kanjun Qiu, "Generally Intelligent #0: A podcast for deep learning researchers", Imbue, 2020-12-06. imbue.com/...2020-12-06-podcast-episode-0-intro. Accessed 2026-10-27.
  12. ^1 ^2Joshua Albrecht, Abraham J. Fetterman, Bryden Fogelman, Ellie Kitanidis, Bartosz Wróblewski, Nicole Seo, Michael Rosenthal, Maksis Knutins, Zachary Polizzi, James B. Simon, Kanjun Qiu, "Avalon: A Benchmark for RL Generalization Using Procedurally Generated Worlds", arXiv:2210.13417, 2022-10-24. arxiv.org/...2210.13417. Accessed 2026-10-27.
  13. ^NeurIPS, "Avalon: A Benchmark for RL Generalization Using Procedurally Generated Worlds (Datasets and Benchmarks Track)", NeurIPS Proceedings, 2022-12. proceedings.neurips.cc/...-Datasets_and_Benchmarks. Accessed 2026-10-27.
  14. ^1 ^2 ^3 ^4 ^5Maginative, "AI Startup Imbue Raises Additional $12 Million, Extending Series B to Over $210 Million", Maginative, 2023-10-23. maginative.com/...g-series-b-to-over-210-million-2. Accessed 2026-10-27.
  15. ^1 ^2 ^3Latent Space, "State of the Art: Training >70B LLMs on 10,000 H100 clusters (with Josh Albrecht and Jonathan Frankle)", Latent Space, 2024-06-25. latent.space/...llm-training-2024. Accessed 2026-10-27.
  16. ^Imbue, "Training >70B LLMs on 10,000 H100 clusters: Josh on the Latent Space podcast", Imbue Talks, 2024-06-25. imbue.com/...latent-space-70b. Accessed 2026-10-27.
  17. ^1 ^2imbue-ai, "sculptor (GitHub repository)", GitHub. github.com/...sculptor. Accessed 2026-10-27.
  18. ^1 ^2 ^3Alessio Fanelli and Shawn (swyx) Wang, "Why AI Agents Don't Work (yet), with Kanjun Qiu of Imbue", Latent Space, 2023-10-14. latent.space/...imbue. Accessed 2026-10-27.
  19. ^1 ^2 ^3Imbue, "Open-sourcing CARBS: how we used our hyperparameter optimizer to scale up to a 70B-parameter language model", Imbue Research, 2024-06-25. imbue.com/...70b-carbs. Accessed 2026-10-27.
  20. ^1 ^2imbue-ai, "cluster-health (GitHub repository)", GitHub, 2024. github.com/...cluster-health. Accessed 2026-10-27.
  21. ^imbue-ai, "carbs (GitHub repository)", GitHub. github.com/...carbs. Accessed 2026-10-27.
  22. ^1 ^2 ^3Imbue, "Scaling Laws For Every Hyperparameter Via Cost-Aware HPO", Imbue Research, 2023-06-10. imbue.com/...carbs. Accessed 2026-10-27.
  23. ^1 ^2 ^3 ^4Imbue, "We build tools that put software under your control (About Imbue)", Imbue. imbue.com/about. Accessed 2026-10-27.
  24. ^1 ^2 ^3 ^4Kanjun Qiu, "Kanjun Qiu (personal website)". kanjun.me. Accessed 2026-10-27.
  25. ^Outset Capital, "Outset Capital: Builders Backing Builders (home page)". outsetcapital.com. Accessed 2026-10-27.
  26. ^1 ^2 ^3 ^4 ^5 ^6Generally Intelligent, "Introducing Generally Intelligent", Imbue Blog, 2022-10-20. imbue.com/...launch. Accessed 2026-10-27.
  27. ^1 ^2 ^3 ^4 ^5 ^6Imbue, "Imbue raises $200M to build AI systems that can reason and code", Imbue Blog, 2023-09-07. imbue.com/...introducing-imbue. Accessed 2026-10-27.
  28. ^1 ^2 ^3 ^4 ^5Imbue, "Imbue Raises Fresh Capital, Amazon's Alexa Fund Joins Investor Lineup", Imbue Blog, 2023-10-19. imbue.com/...imbue-raises-fresh-capital. Accessed 2026-10-27.
  29. ^Imbue, "Sculptor: Catch and fix issues as you code", Imbue Blog, 2025-04-08. imbue.com/...sculptor. Accessed 2026-10-27.
  30. ^Josh Albrecht (thejash), "Show HN: Sculptor - A UI for Claude Code", Hacker News, 2025-09-30. news.ycombinator.com/item. Accessed 2026-10-27.
  31. ^1 ^2 ^3 ^4Josh Albrecht, "A more radical Imbue", Imbue Blog, 2026-01-30. imbue.com/...a-more-radical-imbue. Accessed 2026-10-27.
  32. ^1 ^2Danver Braganza, "Offload: We made agents run our test suite 6x faster", Imbue Blog, 2026-03-17. imbue.com/...offload. Accessed 2026-10-27.
  33. ^1 ^2Imbue, "Sculptor: Extensible, parallel coding workspace you own (product page)", Imbue. imbue.com/...sculptor. Accessed 2026-10-27.
  34. ^Imbue, "Usefully run hundreds of Claudes in parallel with mngr", Imbue Blog, 2026-04-02; and "mngr (product page)", imbue.com/...mngr. imbue.com/...mngr. Accessed 2026-10-27.
  35. ^Andrew Laack, "Vet: Prevent coding agents from making mistakes", Imbue Blog, 2026-03-05. imbue.com/...vet. Accessed 2026-10-27.
  36. ^Hynek Urban, "Latchkey: Use plain old REST instead of MCP", Imbue Blog, 2026-02-27. imbue.com/...latchkey. Accessed 2026-10-27.
  37. ^Nayana Bannur and Nikos Plugachev, "Blueprint: One-shot bigger coding tasks", Imbue Blog, 2026-04-28. imbue.com/...blueprint. Accessed 2026-10-27.
  38. ^Thad Hughes and Catherine Kim, "Introducing Keystone: Building self-configuring agents", Imbue Blog, 2026-03-25. imbue.com/...keystone. Accessed 2026-10-27.
  39. ^Millan Philipose, "Bouncer: Heal your feed", Imbue Blog, 2026-04-08. imbue.com/...bouncer. Accessed 2026-10-27.
  40. ^Darren Jia and Millan Philipose, "Bouncer 2.0: Leveraging Local Compute to Detect AI Slop", Imbue Blog, 2026-07-17. imbue.com/...aging-local-compute-to-detect-ai-slop. Accessed 2026-10-27.
  41. ^Imbue, "Cloud in a Bottle (product page)", Imbue. imbue.com/...cloud-in-a-bottle. Accessed 2026-10-27.
  42. ^1 ^2 ^3 ^4 ^5 ^6Josh Albrecht, "Introducing Imbue Studio: the future of personal computing", Imbue Blog, 2026-10-01. imbue.com/...studio-announce. Accessed 2026-10-27.
  43. ^Josh Albrecht, "It's time for Personal Computing 2.0", Imbue Blog, 2026-10-01. imbue.com/...it-s-time-for-personal-computing-2-0. Accessed 2026-10-27.
  44. ^1 ^2Daniel Mewes, Nayana Bannur and Catherine Kim, "Beating ARC-AGI-2 with Code Evolution", Imbue Blog, 2026-02-27. imbue.com/...2026-02-27-arc-agi-2-evolution. Accessed 2026-10-27.
  45. ^1 ^2Daniel Mewes, Nayana Bannur and Catherine Kim, "LLM-based Evolution as a Universal Optimizer", Imbue Blog, 2026-02-27. imbue.com/...2026-02-27-darwinian-evolver. Accessed 2026-10-27.
  46. ^1 ^2 ^3Daniel Mewes, "Automating AI model research with evolution", Imbue Blog, 2026-07-21. imbue.com/...2026-07-20-imbue-catalyst-nanochat. Accessed 2026-10-27.
  47. ^Daniel Mewes, "Autonomous theory discovery", Imbue Blog, 2026-07-22. imbue.com/...07-20-imbue-catalyst-theory-discovery. Accessed 2026-10-27.
  48. ^1 ^2 ^3 ^4Kanjun Qiu and Josh Albrecht, "There Will Be a Scientific Theory of Deep Learning", Imbue Blog, 2026-04-24. imbue.com/...isode-38-jamie-simon-and-daniel-kunin. Accessed 2026-10-27.
  49. ^1 ^2Jamie Simon, Daniel Kunin, et al., "There Will Be a Scientific Theory of Deep Learning", arXiv:2604.21691, 2026-04-23. arxiv.org/...2604.21691. Accessed 2026-10-27.
  50. ^Matt Boulos, "A Healthy Ecosystem for AI Agents: Introducing the AFI", Imbue Blog, 2025-11-07. imbue.com/...tem-for-ai-agents-introducing-the-afi. Accessed 2026-10-27.
  51. ^1 ^2 ^3Reed Albergotti, "AI agents are getting a lobby group in DC", Semafor (via Yahoo News), 2025-10-29. yahoo.com/...ai-agents-getting-lobby-group-175036465. Accessed 2026-10-27.
  52. ^Imbue, "Closed vs. open AI ecosystems: Kanjun on Decoder", Imbue Blog, 2025-10-09. imbue.com/...-open-ai-ecosystems-kanjun-on-decoder. Accessed 2026-10-27.
  53. ^Abraham J. Fetterman, Ellie Kitanidis, Joshua Albrecht, Zachary Polizzi, Bryden Fogelman, Maksis Knutins, Bartosz Wróblewski, James B. Simon, Kanjun Qiu, "Tune As You Scale: Hyperparameter Optimization For Compute Efficient Training", arXiv:2306.08055, 2023-06. arxiv.org/...2306.08055. Accessed 2026-10-27.
  54. ^1 ^2Imbue, "Imbue Blog (index of posts, podcast episodes and press)", Imbue. imbue.com/blog. Accessed 2026-10-27.
  55. ^Imbue, "Website Terms of Use", Imbue. imbue.com/...terms. Accessed 2026-10-27.

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

8 revisions by 1 contributor · v8 · 5,386 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Funding rounds, valuation, products, research releases and 2026 reorganization checked against Imbue's own posts, TechCrunch, Crunchbase News, Dell and Semafor coverage, October 2026.

Cite this page: AI Wiki. "Imbue." aiwiki.ai, updated 28 Oct 2026, fact-checked 27 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/imbue_ai

Edit article