Preference Model
AI NativePreference Model develops reinforcement-learning environments for training and evaluating AI models, and publishes Karotte, an open-source framework for building and running those environments.
Est 2025
Preference Model develops reinforcement-learning environments for training and evaluating AI models, and publishes Karotte, an open-source framework for building and running those environments.
Published Oct 13, 2026
| Starting price | Commercial environments: unpublished; Karotte software: free | SDK/platform access free; published per-environment hourly pricing on the vendor homepage; enterprise pricing by quote | Verifiers open-source; sandboxes: published per-resource infrastructure pricing | No public starting price identified; demo-led sales |
| Suitable for | Specialist ML-engineering environments and buyer-operated hardened tasks | Hosted environment development, evaluation, traces, and task QA | Integrated open-model development and scalable sandbox execution | Broad professional, computer-use, and enterprise-tool simulations |
| Key strength | Opinionated submission/grading protections and inspectable implementation | Integrated platform, QA agents, debugging, and cloud execution | Environment library, native sandbox integration, and wider training stack | Enterprise workflow coverage and expert-curated artifacts |
| Key weakness | Commercial terms/assurance unclear; cloud integration left to buyer | Cloud charges and feature eligibility need confirmation | Infrastructure pricing does not include the whole model-development cost | Opaque pricing; environment fidelity and gains require task-specific validation |
| Deployment orientation | Self-contained environment images; local runner | Local SDK plus hosted platform | Open framework plus hosted services | Commercial environments compatible with existing execution pipelines |
—
HUD’s public pages show differing published per-environment hourly rates across its materials. Use a written quote to resolve the discrepancy. Its standalone pricing URL was attempted but inaccessible; the accessible homepage contains pricing. (Vendor product overview and feature pages)
Preference Model’s October 2026 framework comparison tests pinned default configurations against selected scripted attacks. It provides specific test details, but remains vendor-conducted. Findings should not be generalized to every configuration, subsequent release, managed service, or adversarial workload. (Vendor blog / technical deep dive)
No reliable market-share estimate or independently measured customer adoption trend was found. Karotte’s repository displayed a small number of stars and forks in the retrieved snapshot; these indicate early attention, not paid adoption. (Official project repository)
The wider market is expanding: a competitor reported that a significant portion of its new data-training projects involved RL environments. That is a vendor-reported internal metric, with no independent methodology verification, and does not establish Preference Model’s share. (Vendor blog post)
No verified customer testimonials directly comparing Preference Model with these competitors were found.