Imbue

Imbue develops AI systems that reason and code, with the goal of making agents capable of accomplishing larger tasks for their users. Formerly Generally Intelligent, the company announced its rename alongside a $200 million Series B in September 2023. Its work has expanded from reinforcement-learning environments and foundation-model training to Sculptor, a parallel coding-agent workspace, and open tools for configuring software environments and automating computational research.[3][2][9][11][12]

The reasoning-model program followed earlier work on recruiting systems and embodied agents. Those projects supplied concrete lessons about user trust, planning, and the ability to inspect an agent’s decisions.[1]

Co-founders Kanjun Qiu and Josh Albrecht had previously worked together on Ember, which built a laser-raster-projector VR headset, and Sorceress, an AI recruiting company. In her October 2023 interview, Qiu recalled that Sorceress identified candidates and helped employers contact them, using human writers alongside automated email customization. She said the difficulty of showing users why candidates had been selected informed Imbue’s work on trustworthy agent interfaces. Qiu dated serious exploration of the new AI work to late 2019 and early 2020.[1][10]

Generally Intelligent introduced itself publicly on October 20, 2022 as an independent research company developing generally capable AI agents. It explicitly excluded medium-term commercialization milestones from its project. The announcement reported more than $20 million contributed by backers and more than $100 million in future funding commitments tied to options and technical milestones. Qiu separately recalled raising a $20 million Series A in late 2021.[2][1]

At the public launch, the company open-sourced Avalon, a reinforcement-learning simulator and benchmark built on Godot. The Avalon paper, submitted four days later and accepted to NeurIPS Datasets and Benchmarks 2022, described 20 tasks in procedurally generated 3D worlds, including eating, throwing, hunting, and navigation. Tasks shared a reward function, world dynamics, and action space; changing the environment tested generalization within tasks, between tasks, and to tasks requiring combinations of previously learned skills. Joshua Albrecht and ten co-authors developed the benchmark.[2][6]

Qiu said Avalon’s adjustable difficulty gave agents a curriculum with more frequent reward signals. Agents learned mechanical skills such as climbing and opening doors with multiple switches, but struggled with higher-level tasks and consistent execution. She described these limitations, together with users’ need to control and inspect agents, as reasons to experiment seriously with reasoning in natural language. She credited inspiration from Chelsea Finn’s work, identified in the interview introduction as SayCan.[1]

The September 7, 2023 announcement set out a strategy combining reasoning-specific pretraining data, additional computation during inference, internal agents, interfaces, infrastructure, and learning theory. Imbue treated reasoning as the principal obstacle to effective agents: handling uncertainty, gathering information, revising plans, and deciding when to ask users questions. Coding served as an initial test bed because the team wrote code daily, tests could assess results, and generated code let agents act on a computer. At that stage, Imbue said it did not intend to turn its internal coding agents into production products.[3]

Internal prototypes included agents for fixing type errors, generating tests, and formatting code. Qiu described interfaces that let users inspect execution, fork a run, and modify prompts or plans, while acknowledging that broader tasks remained unreliable. Beyond coding, she said Imbue built an agent to analyze 20,000 pages of AI-policy submissions to the Department of Commerce, summarizing the problems and solutions proposed.[1]

Imbue announced its $200 million Series B at a valuation above $1 billion, with participation from Astera Institute, NVIDIA, Kyle Vogt, Simon Last, and others. On October 19, 2023, it announced another $12 million from Amazon’s Alexa Fund and Eric Schmidt, extending the previous fundraise to more than $210 million.[3][4]

On November 28, 2023, Dell and Imbue announced a $150 million infrastructure agreement for a Dell PowerEdge cluster with nearly 10,000 NVIDIA H100 GPUs. Dell said Imbue was already using the system for model training and prototype agents. The companies designed smaller experimental clusters that could network into a large training cluster; Voltage Park managed the system. The agreement was separate from Imbue’s equity financing.[5]

In June 2024, Imbue described a 70-billion-parameter model pretrained from scratch on two trillion tokens. It reported that, after fine-tuning on multiple-choice reasoning benchmarks, the model outperformed untuned GPT-4o evaluated zero-shot on those benchmarks; neither evaluation used chain-of-thought. About a dozen engineers and researchers completed the project, including infrastructure, evaluations, and hyperparameter optimization. This was project staffing, rather than a company-wide headcount.[7]

Imbue credited CARBS, its cost-aware hyperparameter optimizer, with scaling the training system to 70B parameters on the first attempt with minimal instability and no loss spikes, following thousands of smaller-model experiments. Qiu had attributed CARBS’ origin to a colleague named Abe automating his research process and drawing on plasma physics for local search. The June release shared CARBS, cleaned and extended versions of 11 NLP benchmarks, the CodeComprehension benchmark, 450,000 human judgments about question quality, and infrastructure scripts. A released Llama 3-based question-quality model was distinct from Imbue’s from-scratch model.[7][1]

The accompanying infrastructure account focused on a 4,088-H100 cluster across 511 computers and credited Voltage Park, Dell, H5, and NVIDIA. Engineers rewired an initially disconnected InfiniBand fabric, used deterministic data loading to reproduce crashes caused by malformed training data, and modified NCCL logging to identify machines blocking collective operations. Automated recovery reran health checks and resumed from checkpoints on healthy hosts. Coordinating garbage collection across workers eliminated reported throughput sags. The shared tooling excluded a Dell-authored PCIe-switch-bus diagnostic script.[8]

Sculptor began as an April 2025 research preview for catching and fixing issues in agent-written code. In its September 26 announcement, Qiu described a rebuild around community requests for parallel agents, persistent context, and quick verification. Each agent ran in its own container. Pairing Mode synchronized an agent’s files and Git state with a local repository in both directions, enabling IDE testing and direct collaboration. Sessions retained plans, chats, tool calls, and code changes; the interface flagged potential merge conflicts and could return them to an agent for resolution. A December 3 update added Codex support.[9]

In January 2026, Albrecht recalled shipping Sculptor in early October and described stress and burnout during the push to build and stabilize it. He proposed moving toward “more, smaller open software projects” and giving project owners authority to make or delegate decisions. The retrospective framed this as a return to the founders’ emphasis on human agency, including how Imbue itself operated.[10]

Released in March 2026, Keystone is an open-source Python CLI that drives a coding agent to create Dockerfiles, dev containers, and test entrypoints for repositories. Thad Hughes and Catherine Kim described running agents such as Claude Code or Codex inside a Modal sandbox with its own Docker daemon, crediting Modal for the nested execution capability. The agent is prompted to make as many tests as possible pass and provide JUnit XML reports as execution evidence; complex or legacy projects can require difficult environment configuration.[11]

In July 2026, Imbue released the experimental Catalyst research tool under AGPL-3.0. Daniel Mewes described an evolution-inspired system whose central objects are “interpretation strands”: agents’ interpretations of experiments and literature, paired with research roadmaps. Catalyst maintains a population of these strands, shares experimental results, periodically integrates observations into theories, branches alternative interpretations, and scores strands using candidate-solution performance and roadmap novelty.[12]

Using Andrej Karpathy’s AutoResearch setup, Catalyst optimized nanochat candidates within a five-minute training budget on one H100 GPU. Imbue reported a validation bits-per-byte score of 0.9361 after 340 experiments and said improvement had not plateaued when it stopped the run. A linear Catalyst comparison disabled evolution while keeping the base model, hardware, and prompts identical. Comparisons with Recursive’s reported H100 results used estimated experiment counts, leaving relative experiment efficiency uncertain.[12]