EnvLoop / Use cases

Any company can post-train models on its own data and environments.

EnvLoop is built for companies that want a model to get better at their own work. Each use case says what EnvLoop does today and what is arranged with our team.

Build your decision model Run the RL loop

Build your own Jev-class decision model

Post-train a decision model that answers typed questions about your users or business with a calibrated probability from a single forward pass. EnvLoop’s released Pev-27B reaches 0.915 family-macro accuracy on the Pev-Bench public test set, ahead of TypeSafe’s Jev at 0.788 and its base model at 0.762. Training your own is experimental and arranged with our team.

Build your decision model

Run the RL loop on your own environments

Agents attempt your tasks in sandboxed environments, verifiers score the attempts, the verified attempts become training data, and a LoRA SFT Train Job post-trains the model before it is re-evaluated against its base. Rewards drive evaluation and data selection; the training step is supervised fine-tuning.

Run the RL loop

Improve a model round by round with Model Loops

A Model Loop runs recursive self-improvement rounds on one task: it trains candidates, evaluates them on a benchmark and reports whether a candidate beats the base model, with scores by version, audits and spend. You follow it under RSI LOOPS → Model in the Console and start or steer it through the REST API, MCP or the preview CLI.

Read about Model Loops

Post-train and serve a model on your business data

Submit LoRA SFT Train Jobs on Qwen3.8-27B or GLM-5.3-Flash with your own records, download the checkpoints, and serve a verified checkpoint from an OpenAI-compatible endpoint that scales to zero when idle. DeepSeek-V4.1-Flash-1017 is a batch training preview for enabled projects, without a serving profile.

Read about Train Jobs

Evaluate agents with verifiers and publish benchmarks

Run an agent against an Environment Dataset as a Job. Each Task runs in a sandbox, an independent verifier scores the attempt, and the Job records rewards, score coverage and pass metrics for comparison. The catalog includes terminal tasks such as envloop/terminal-bench-2-1-lite, and cu-bench runs this way as the official envloop/cu-bench Runner; Pev-Bench ships its tasks with a Harbor verifier.

Read about evaluation

Evaluate computer-use agents on desktop software

cu-bench tests agents on professional work in 20 Linux desktop applications from screenshots, mouse and keyboard, graded by file-state verifiers. On EnvLoop it runs as the official envloop/cu-bench Runner over private datasets, with access arranged by our team. On the frozen release of 2026-10-16, Claude Opus 5.5 reached 83.5% strict success and gpt-6-astra 22.2%.

View the cu-bench leaderboard

Get environments, verifiers and data for your own training

The task environments, verifiers, benchmarks and training data that EnvLoop builds can be delivered per project or licensed on their own, across desktop software, terminals, operations and finance. Availability and access are arranged with our team.

Explore Off-the-Shelf

Plan a private deployment with our team

Companies with strict data boundaries can plan a private deployment, dedicated support and service commitments with our team. Scope and availability are agreed for each deployment.

Talk about Enterprise