Prime Intellect's Model Factory Merges AI Training and Deployment Into One Loop Prime Intellect launched Model Factory, a managed platform that combines pre-training, supervised fine-tuning, reinforcement learning, inference, sandboxed agent execution and production trace collection in one dashboard and CLI, replacing its earlier Lab platform. Hosted Training is in beta, requires an access request and is billed per GPU-hour prorated to the second, while dedicated deployments add reserved capacity with direct Slurm and Kubernetes access on customer or Prime Intellect GPUs. Prime Intellect reported that its Prime Mega MoE kernels reached 48% MFU on H200s training Qwen's 235B MoE, and that RL runs scale to tens of thousands of isolated microVM sandboxes; checkpointing is off by default and shared LoRA adapter serving has been retired. Prime Intellect's Model Factory Merges AI Training and Deployment Into One Loop Prime Intellect's Model Factory puts pre-training, fine-tuning, large-scale RL, sandboxes, inference and trace collection on one stack, so teams can own their models' full improvement loop. - Prime Intellect launched Model Factory https://www.primeintellect.ai/blog/introducing-model-factory?utm source=alphasignal , which combines pre-training, SFT, RL, inference and trace collection on one platform. - On-demand Hosted Training https://docs.primeintellect.ai/hosted-training/getting-started?utm source=alphasignal is in beta, needs an access request, and is billed per GPU-hour, prorated to the second. - Dedicated deployments add reserved capacity plus direct Slurm and Kubernetes access, on your GPUs or Prime Intellect's. - Prime Mega MoE https://www.primeintellect.ai/blog/prime-mega-moe?utm source=alphasignal kernels hit 48% MFU on H200s training Qwen's 235B MoE. - RL runs scale to tens of thousands of isolated microVM sandboxes, and Prime Traces turns production failures into new training environments. - Watch out: checkpointing is off by default, and shared LoRA adapter serving has been retired. AI infrastructure company Prime Intellect has launched Model Factory, a managed platform for training, deploying, observing, and retraining open-weights models. It combines pre-training, supervised fine-tuning, reinforcement learning, inference, sandboxed agent execution, and production traces through one dashboard and CLI. The product replaces Prime Intellect’s earlier Lab platform and is available through beta on-demand jobs or dedicated deployments. A closed-model API limits how developers can adapt and operate a model: the provider receives prompts, controls available versions, and may prevent training against a product’s own reward signal. Running open weights on controlled infrastructure gives teams more authority over data paths, model versions, and training. Model Factory packages the clusters, training framework, inference servers, and sandboxes required to do that. On-demand workloads run on Prime-operated GPUs, while dedicated deployments can use customer hardware. One control plane for the whole loop Model Factory supports four training stages described in the launch post https://www.primeintellect.ai/blog/introducing-model-factory?utm source=alphasignal . Pre-training learns general patterns from a large corpus, mid-training continues that process on targeted data, supervised fine-tuning SFT learns from labeled examples, and reinforcement learning RL updates weights from scored attempts. 1. Train. Choose an open model or bring your own, then define tasks and an environment. The environment is the sandbox in which an agent works. It supplies tools and context, then applies a reward function to each attempt. Training runs on Prime Intellect’s open-source prime-rl framework, with integrations for its verifiers library, the community Environments Hub, and Harbor benchmark format. 2. Deploy. Launch an inference endpoint from the dashboard. Training and inference can run on the same hardware, and the endpoint uses the same Prime Inference API offered for GLM, DeepSeek, and Kimi models. 3. Observe and improve. Prime Traces records tool calls, task paths, outputs, and failures. Teams can turn recurring failures into new environments or examples for the next training run. During an RL iteration, the model generates task attempts and the environment scores them. A coordinator collects those attempts into batches, and the trainer uses the scores to update the model’s weights. Each iteration requires inference capacity, isolated sandboxes for code and tool calls, and environments that can grade results. Model Factory coordinates those systems while researchers specify the tasks, rewards, and training recipe. The machinery under the dashboard - Asynchronous RL. The prime-rl framework keeps generating attempts while the trainer updates weights, reducing idle time between inference and training. Prime Intellect used the stack for its INTELLECT models, including INTELLECT-3, a 106-billion-parameter mixture-of-experts model with 12 billion active parameters. - Hardware-isolated sandboxes. Agents execute inside microVMs that can be checkpointed, copied, and restricted. Prime Intellect says a training run can use tens of thousands of concurrent sandboxes. - Faster mixture-of-experts training. A mixture-of-experts model routes each token to a small set of specialist subnetworks, often spread across several GPUs. Moving those tokens between devices can become a bottleneck. Prime Intellect added a training backward pass to DeepSeek’s fused Mega MoE kernel, a GPU routine that combines operations to reduce overhead. The implementation overlaps computation with communication. Prime Intellect reports that its Mega MoE kernel https://www.primeintellect.ai/blog/prime-mega-moe?utm source=alphasignal delivered up to 1.34 times the end-to-end throughput of its previous prime-rl implementation. Using Qwen’s 235-billion-parameter MoE architecture on Nvidia H200 GPUs, it reached 48% model FLOPs utilization, a measure of how much theoretical GPU arithmetic capacity performs model computation. Two ways to buy the stack | Mode | How it runs | Typical use | Access and billing | |---|---|---|---| | On-demand Hosted Training | Prime Intellect runs a submitted configuration on its GPU fleet. | Managed jobs using supported frameworks and training recipes. | Beta approval required; billed per GPU-hour and prorated per second. | | Dedicated Model Factory | Reserved capacity runs on customer or Prime Intellect hardware, with managed training and inference plus direct Slurm and Kubernetes access. | Custom architectures, framework changes, and specialized rollout strategies. | Reserved capacity; public rates are not listed. | On-demand rates vary by GPU type and appear in the CLI. There is no minimum runtime charge, and Prime Intellect does not bill for queueing, scheduling, or image pulls. Persistent volumes cost $0.09 per GiB per month. Trace ingestion and storage carry separate fees. The CLI flow below installs the client, authenticates, lists available models and rates, creates persistent storage, starts a run, and follows its logs: uv tool install -U prime prime login prime train models prime volumes create research --size 100Gi prime train configs/rl.toml --volume research prime train logs