AI21 Lab-as-a-Service

According to the page, AI21's Forward-Deployed Researchers (FDRs) train specialized models for a customer's task, positioned as creating an AI edge competitors cannot buy. The service covers evaluation suite building from real tasks, training recipe development using on- and off-policy methods, joint tuning of the model and harness, and optimizing models for production serving, managed by AI21 or deployed in the customer's environment. Training, evals and serving can run inside the customer's VPC or on-premise, with scoped, logged access and no data retention by AI21 as stated. The page also cites partner results, including post-training gains and matching or exceeding frontier model quality on specific tasks, and references a Platform with Evaluate, Train, Serve and Improve stages.

Key features

• Forward-Deployed Researchers who post-train specialized models for a specific task
• Model and harness trained together, with a harness optimizer for agent configuration
• Evaluation suite built from real tasks, left with the customer at engagement end
• Gating of quality, cost and latency throughout training
• Model optimization for fast and efficient serving
• Managed or embedded deployment, on customer infrastructure or AI21's
• Training, evals and serving can run in the customer's VPC or on-premise
• Scoped, temporary, logged and auditable expert access
• No retention of customer data copies and full output traceability

Best suited for

• Enterprises wanting specialized models trained on their proprietary data and workflows
• Organizations requiring data to stay inside their own VPC or on-premise environment
• Teams building specialized AI agents that need post-training on real production tasks

Pricing

Not stated on the website