The Enterprise Case for Open-Weight Models on Databricks

Enterprises are rethinking reliance on single frontier model providers due to rising costs, shifting model rankings, and restrictive guardrails that block legitimate tasks. Open-weight models now match frontier performance at lower expense, prompting companies to reassess their AI architectures. Data retention policy changes and demand for governance visibility further complicate single-provider strategies. A flexible approach is needed to incorporate innovations without sacrificing operational freedom or productivity gains.

Unity Gateway enables centralized governance while supporting both open-weight and frontier models through a unified API. Customers have increased their average number of unique models from 4.2 to 7.1 since January, with roughly 40% of those being open-weight. The platform allows model switching via configuration, avoiding code changes, and supports portable harnesses like Codex and OpenCode. This flexibility aligns with growing enterprise demand for adaptable AI deployments.

Databricks optimizes open-weight model performance using tailored inference engines and frontier GPUs to maintain cost efficiency at scale. Independent evaluations show GLM 5.3 achieves a 9.1-second end-to-end response time on Databricks, comparable to top-tier managed endpoints. After GLM 5.2’s mid-July release, its usage surged to ~40% of open-weight tokens on Unity Gateway within a month, indicating rapid adoption of competitive open-weight options.

To meet demand for GPU capacity, Databricks raised $3 billion in July, with CEO Ali Ghodsi citing global shortages across regions. Open-weight models hosted on Unity Gateway now integrate with existing coding harnesses via OpenAI-compatible endpoints, requiring only configuration changes. Databricks emphasizes a portfolio approach, balancing open-weight models for cost and control with frontier models for superior results, all managed through unified governance and routing policies.