The Enterprise Case for Open-Weight Models on Databricks

As field engineers at Databricks, we spend our days inside enterprise AI deployments, where we’ve observed a clear shift in strategy. After two years of standardizing on single frontier model providers, companies are reevaluating their approach due to seven key pressures:

1. Spend is not generating sufficient returns. AI budgets have outpaced revenue and cost savings, prompting a need for better unit economics [1].

2. Model rankings have changed repeatedly since 2023. The best model for coding, reasoning, and extraction changes month to month, and the top intelligence ranking has changed hands 20+ times since May 2023 [2]. Architectures tied to a single provider make it harder to adopt a better model as rankings change.

3. Frontier guardrails can block legitimate work. Model guardrails imposed by AI labs can refuse legitimate tasks, such as analyzing live malware or reproducing an exploit to patch it, a limitation discovered in the recent Hugging Face incident [6][7]. When Hugging Face responders fed real attack logs and exploit payloads to hosted frontier models for analysis, the guardrails refused, and the team ran the analysis on an open-weight model (GLM 5.2) instead.

4. Data retention policies have evolved and may continue to do so. Some frontier providers have deviated from zero-data-retention (ZDR) policies, such as the 30-day minimum retention introduced with Anthropic's frontier tier [8]. For enterprises whose compliance policy was written around the principle that "no prompt or completion is ever stored," an upstream policy change could introduce problems and potentially require a system change.

5. Companies want visibility into and control over model governance. Demand is growing for end-to-end visibility and governance across AI models. Customers want searchable records of model inputs, outputs, errors, and agent actions to enable teams to evaluate accuracy and improve performance, as well as centralized controls to protect sensitive data and intellectual property.

6. The harness matters as much as the model. Databricks’ research shows that the scaffolding around a model, its harness of prompts, tools, and control flow, changes measured performance. In Databricks’ own benchmark, the same model scored 61.7% in one harness, 48.1% in another, and 38.4% bare [9]. Because harness choice changes benchmark results, evaluations should test the model and harness together.

7. Open-weight models caught up to frontier models. The top open-weight models now benchmark at a level comparable to that of top frontier models at a lower cost [3][4][5]. This was not true eighteen months ago. Open-weight models achieving frontier-level benchmark performance change the cost, control, and deployment trade-offs described in the other six points.

Across these observations, companies are trying to: 1) optimize costs without sacrificing productivity gains, 2) preserve operational freedom, and 3) ensure AI implementations can quickly incorporate innovations. To do so, they need an architecture in which models and harnesses can change without weakening governance.

Commoditized Intelligence: As model intelligence becomes a commodity, open-weight options now provide intelligence parity with top frontier models at a fraction of the cost, enabling better ROI on AI deployments.

Data Sovereignty: Hosting models on controllable infrastructure gives teams authority over compliance, security, and data retention policies and reduces the risk of non-compliant changes in the future.

Complete Operational Autonomy: Open-weight models let teams run workloads that frontier guardrails may inhibit (e.g., cybersecurity response)

Source: [5] DeepSWE Leaderboard by Cost. Some of the top open-weight models compared against their proprietary equivalents, filtered at the authors’ discretion.

Unity Gateway serves open-weight and frontier models through a unified API with centralized governance controls, enabling teams to switch between models via configuration or smart routing [10][11] rather than through code changes or other abstractions [12][13].

The multi-model capability of Unity Gateway is becoming integral to customers’ AI strategies. Since January, the average number of unique models used per customer has increased from 4.2 to 7.1, an indicator of the options available on Unity Catalog and of the pursuit of higher-ROI models. Of those models, roughly ~40% were open-weight.

Source: Databricks’ internal data from the week ending January 5, 2026, through the week ending July 27, 2026. Created using Genie One.

Serving large open-weight models with competitive response times requires optimization to maintain cost efficiency at scale. Databricks addresses latency constraints by leveraging tailored inference engines and optimized kernels deployed on frontier GPUs. Independent evaluations from Artificial Analysis highlight these performance characteristics: for instance, GLM 5.3 (max) achieves an end-to-end response time (including input processing, reasoning, and generation) of 9.1 seconds on Databricks [15], performing favorably alongside top-tier managed endpoints [16].

Source: [15] GLM 5.3 Model Providers: Artificial Analysis

Smart routing sends sessions to the cost-efficient model for the task, so routine tasks shift off of frontier models [10][11].

Fallback chains automatically reroute in response to model rate limits or provider outages [14].

Budgets with spend caps by user, team, or use case prevent runaway spending [19][30]. These limits can be configured with varying levels of severity, from soft warnings to strict policies that reroute sessions to cheaper models.

Centralized governance for tools and skills. Along with built-in MCP servers, Unity Catalog supports registering External MCP servers that run through managed proxies and are audited centrally. Hence, agents get their tools with the same governance as their models [20]. Skills [21] can also be registered and governed, improving determinism and token efficiency for repeatable workflows.

Unified governance across open-weight, external, and frontier models. Connect external model providers (e.g., Amazon Bedrock, Azure AI Foundry) through model provider services [17] and route requests through the same central governance layer used for Databricks-hosted open-weight and frontier models [18]. Manage access, spend controls, and audit trails across your entire model portfolio from a single platform.

GLM 5.2 coincided with a sharp increase in open-weight usage on Unity Gateway. After GLM 5.2 became available in mid-July, its usage rose. At the same time, growth among Unity Gateway customers remained relatively consistent, suggesting that existing customers shifted more AI traffic toward open-weight models. Just one month after first being offered, GLM 5.2 already accounts for ~40% of open-weight tokens used on Unity Gateway, with most of that growth incremental. Both GLM 5.3 and Kimi K3's benchmark performance give customers another open-weight option to evaluate alongside GLM 5.2 [5].

Source: Databricks’ internal data from the week ending January 5, 2026, through the week ending July 27, 2026. Created using Genie One

To support this demand, Databricks announced a $3 billion fundraising round in July, and CEO Ali Ghodsi told CNBC the raise is partially about securing GPU capacity to host open models. "We are running out of GPUs everywhere... Japan, Korea, U.S., and India, so we just needed to get a lot of these, and that requires a lot of funding. So that’s what triggered this fundraiser for us" [22].

The harness your developers already live in can still run open-weight models hosted on Databricks. As of late July, models served on Unity Gateway, including GLM 5.3, can run inside provider coding harnesses such as Codex and OpenCode [12][26], with support for additional harnesses rolling out. The gateway exposes OpenAI-compatible endpoints per harness, so the setup is just a config change. Point the harness at your workspace's gateway URL, or use the ucode helper with one line:

ucode codex --model system.ai.glm-5-3

Enterprises should plan for a portfolio of models because no single model or provider will lead in quality, cost, and policy fit for every workload. Of those to choose from, open-weight models now offer competitive intelligence at lower cost and greater control over data handling and governance policies. In contrast, frontier models remain useful where they deliver better results. Unity Gateway manages these options through common access controls, routing policies, budgets, and auditing. Omnigent extends that flexibility to the harness, so teams can change models or agent environments without rebuilding workflows.

The usage data in this article shows that customers are already moving in this direction. They are serving more models, directing more traffic to open weights, and using portable harnesses to broaden their options. The durable advantage is the ability to make those changes quickly as model options, quality, pricing, and policy requirements evolve.

[1] Nanda, M., Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025, July). The GenAI divide: State of AI in business 2025 (Preliminary Findings from AI Implementation Research). Project NANDA, Massachusetts Institute of Technology. [2] LLM Leaderboard History: AI Rankings 2023-2026 Tracked, BenchLM [3] Kimi K3 - Intelligence, Performance & Price Analysis, Artificial Analysis [4] GLM-5.3 (max) - Intelligence, Performance & Price Analysis | Artificial Analysis [5] DeepSWE Leaderboard by Cost [6] Security incident disclosure, July 2026, Hugging Face Blog [7] OpenAI's Hugging Face Breach Shows Frontier AI Guardrails Are Failing, Forbes, July 23, 2026 [8] Statement on the US government directive to suspend access to Fable 5 and Mythos 5 [9] What is an AI Agent Harness?, Databricks Blog [10] Smart routing for coding agents [11] Smart Routing in Unity AI Gateway: Match frontier quality with 30%+ lower cost per task [12] Databricks-hosted foundation models, Databricks docs [13] Agent Bricks: Data + AI Summit 2026, Databricks Blog [14] Configure routing and fallbacks for model services [15] GLM-5.3 (max) API Provider Benchmarking & Analysis [16] https://artificialanalysis.ai/models#end-to-end-response-time [17] Govern external model providers (model provider services) [18] Databricks Unity AI Gateway Documentation [19] Manage budgets for Unity Gateway [20] Unity AI Gateway: How to connect agents to external MCPs securely, Databricks Blog [21] Skills support in Unity Gateway [22] Databricks CEO: We're hosting open source models like Kimi and running out of GPUs, CNBC, July 17, 2026 [23] Models & Credentials, Omnigent docs [24] Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase [25] Governing MCPs at Scale with Unity AI Gateway [26] Use Open Source models with coding agents [27] Unity AI Gateway: Multi-AI governance and cost control, Databricks [28] Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents, Databricks Blog [29] GLM-5.2 Providers: Output Speed Comparison, Artificial Analysis [30] Introducing AI spend controls with Unity Gateway [31] Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents

Learn how to govern, route, and monitor open-weight, frontier, and external models through one unified platform.