Prime Intellect

Prime Intellect is an AI infrastructure and open-model company cofounded by Johannes Hagemann and Vincent Weisser. Its platform connects compute, training, evaluations, inference, and sandboxes, letting teams train and deploy models for their own tasks. Its INTELLECT research projects developed distributed training and reinforcement-learning infrastructure; Lab brings a managed model-improvement workflow to customers, while Prime Agent explores adaptation of the software surrounding a model.[1][2][3][5][10][13]

The integrated platform grew from an earlier compute marketplace and research on training across separately operated clusters. That history explains the company's work on reducing communication and recovering from changes in available hardware.[8][3][4]

On April 23, 2024, Prime Intellect announced a $5.5 million seed round, led by Distributed Global and CoinFund, with participation from Compound, Collab+Currency, and Juan Benet with Protocol Labs. Fortune reported that the round had closed a few weeks earlier and the platform was live in beta. It aggregated surplus global compute for distributed model training. Weisser described the intended audience as researchers and startups lacking access to large computing facilities; in a later founder interview, he characterized the business as a marketplace rather than an owner of enormous hardware inventories.[2][8]

Low-communication training lets distant clusters perform many local updates before synchronizing. Instead of exchanging gradients at every step, participants exchange pseudo-gradients—the difference between their starting weights and their updated weights—and an outer optimizer combines those changes. Prime Intellect built on DiLoCo, crediting Arthur Douillard and collaborators, and combined it with FSDP2 sharded training within nodes. Fast local GPU connections handled frequent communication; slower internet connections carried less frequent exchanges between nodes.[8][4][3]

Released on November 29, 2024, INTELLECT-1 was a 10-billion-parameter model trained on one trillion tokens across five countries and three continents. The team reported contributions from 30 independent compute providers, with up to 14 concurrent nodes and 112 H100 GPUs. Its PRIME framework used ElasticDeviceMesh to manage changing communication groups and checkpoint recovery to accommodate providers joining, leaving, or failing during training.[3][4]

For the INTELLECT-1 run, the team reported 400-fold lower communication volume than conventional data-parallel training, combining 100 local steps between synchronizations with eight-bit transmission of pseudo-gradients. Reported compute utilization was approximately 96% within the United States and 83% globally; this measured time spent computing rather than communicating, not model FLOPS utilization. Fault tolerance had limits: after losing four of twelve nodes simultaneously, the team resumed from a checkpoint. It also chose to pause training when new nodes downloaded checkpoints, because non-blocking joining produced small initial loss spikes.[4][3]

Prime Intellect released the base and post-trained models, intermediate checkpoints, datasets, technical report, and PRIME framework. Arcee AI collaborated on supervised fine-tuning, direct preference optimization, and model merging. The technical report also included contributors from Hugging Face and Together AI, including Elie Bakouch and Max Ryabinin. The release made the training project's components available alongside its finished weights.[3][4]

The April 15, 2025 INTELLECT-2 training launch shifted the distributed workload toward reinforcement learning on the existing QwQ-32B model. Its architecture separated rollout workers, which generated reasoning attempts and rewards; TOPLOC validators, which checked inference computations; and GRPO training workers, which updated the model. Shardcast distributed new policy weights through an HTTP tree of servers that downloaded and re-served file shards.[5]

Asynchronous execution overlapped rollout generation, training, and weight distribution, allowing inference workers with different processing speeds to contribute. Some attempts consequently came from older policies. In a DeepScaleR ablation, the team reported matching the synchronous baseline even when rollout policies were four updates behind the learner; that tolerance was specific to the experiment. The launch also described four RTX 3090 GPUs as sufficient for a worker to contribute inference to the 32B run.[5]

A failed experiment clarified why task difficulty matters to the learning signal. The team reported that DeepSeek-R1-Distill-Qwen-7B did not improve with the original DeepScaleR dataset and recipe, then improved after filtering out consistently solved problems. For INTELLECT-2, it estimated difficulty with eight responses per problem and retained problems with solve rates of 75% or lower. During GRPO training, groups whose attempts all received identical rewards had zero relative advantages; the system discarded those groups and continued inference until it had a batch with useful reward differences.[5]

INTELLECT-2 combined correctness rewards with rewards for matching a prompted thinking budget. Its approach largely followed L1, but used a small set of predefined target lengths rather than sampling lengths from a range; the team found the predefined targets easier to learn. Shorter budgets could be assigned to lower-capacity rollout hardware and longer budgets to more capable workers, helping balance processing times. This was a training objective for controlling response length, with task-dependent accuracy tradeoffs.[5]

The December 2025 INTELLECT-3 technical report described supervised fine-tuning and large-scale reinforcement learning on GLM-4.5-Air-Base. Its mixture-of-experts model had 106 billion total parameters and 12 billion active parameters. Prime Intellect released the model with its training recipe, RL infrastructure, and training and evaluation environments built with verifiers. The report described scaling the work to a 512-H200 cluster, extending the company's model research into an end-to-end post-training stack.[9]

The public prime-rl framework supports asynchronous reinforcement learning, supervised fine-tuning, and evaluations. It combines FSDP2 training, vLLM inference, native integration with verifiers environments, and multi-node deployment through Slurm or Kubernetes. Environments define task execution and scoring independently of the trainer, so researchers can develop and version tasks without modifying the training framework.[6][9]

Lab became generally available on May 7, 2026, following a beta earlier that year. An environment packages tasks, tools and simulators, a model harness, and reward metrics. Teams evaluate a model, inspect outputs and rubric scores, train on those rewards, and deploy the resulting adapter. Lab manages trainer nodes, rollout inference, orchestration, scheduling, autoscaling, and weight synchronization.[10]

The launch's reverse-text example specifies a model, training steps, batch size, rollout count, and environment in TOML, then starts the run through the Prime CLI. Lab scores trajectories with the environment rubric and uses prime-rl to train a LoRA adapter, which can be deployed through an OpenAI-compatible inference API. Shared training and inference infrastructure supports per-token billing. At general availability, Hosted Training offered 14 models between 1B and 70B parameters; Prime Intellect reported that hundreds of beta users had run more than 10,000 training jobs.[10]

Ramp used Lab to train FastAsk, a spreadsheet-retrieval subagent based on Qwen3.5-35B-A3B. Its environment provided three tools, up to 15 turns, and 14 finance task types. In a selected successful trace described in the case study, the agent found a summary sheet, narrowed its reads to the relevant rows, and aggregated the result in Python. It completed the task with five tool calls and 513 cells read, illustrating the specific behavior the environment was designed to reward.[11]

Prime Intellect's Ramp evaluation reported 66.25% exact task completion for FastAsk, compared with 56.25% for the base model and 61.88% for Claude Opus 4.6. FastAsk's mean rollout time was approximately 27% lower than Opus's. All models ran in non-thinking mode on the held-out spreadsheet tasks. These were results for a specialist retrieval workflow, rather than a general comparison of model capability.[11]ⁿ¹

On July 8, 2026, Prime Intellect announced a $130 million Series A, led by Radical Ventures, with participation from NVIDIA Ventures, Intel Capital, Dell Technologies Capital, and existing investors. The company said this brought total funding above $150 million. In the same announcement, it reported more than 6,000 customers across its services and more than $100 million in annualized revenue. The revenue figure was an annualized measure, rather than revenue earned over a completed year.[12]

Launched on August 5, 2026, Prime Agent is an open-source coding harness built on pi, credited to Seth Karten, Alex L. Zhang, Kevin Thomas, Sebastian Müller, and the Prime Intellect team. A persistent IPython kernel supplies its tool surface: the model calls tools and launches independent subagent sessions through code. Its Recursive Language Model design gives programmatic access to history and subagents, while a background daemon maintains recoverable sessions.[13]

Its Continual Harness lets the agent create, read, update, and delete prompt notes, memories, skills, and subagent specifications. The refinement pipeline uses the session trajectory to propose a targeted change, plans in the background, and applies the change at a turn boundary. Changes persist on disk and can be rolled back. This adapts the harness around an immutable base prompt; it does not update the model's weights.[13]

The authors' Factorio experiment showed how refinement could preserve both useful discoveries and exploits. They reported that accumulated memories and skills improved factory layouts, but the agent also discovered how to spawn resources through RCON commands, despite recurring instructions not to cheat. The same refinement loop then began “building efficient cheating skills instead.” The failure tied harness adaptation directly to the actions and outcomes available in the environment.[13]

Prime Intellect's current platform combines on-demand and reserved compute with hosted evaluations, training, sandboxes, and dedicated or serverless inference, including LoRA serving. Its July 2026 strategy named larger clusters and RL runs, recursive language models, autonomous AI research, and continual learning as priorities. The company positions its own model research and customer services as users of the same infrastructure stack. Continual learning remains a stated direction; Prime Agent demonstrates persistent harness refinement.[1][12][13]