Inkling on Databricks: A New Default for Enterprise AI
Inkling is a 975-billion-parameter open-weights AI model from Thinking Machines Lab that enterprises can download, inspect, modify, and now deploy directly on Databricks to power coding workflows and AI agent development without relying on per-token APIs or a single vendor’s infrastructure. This launch matters because it turns open-weights AI from a side project into a credible default for enterprise LLM deployment. Databricks is a day zero launch partner, making Inkling immediately available for teams to apply to their own data and workflows. The message is clear: if you are still building all your AI strategy around closed black-box APIs, you are behind the curve. The Inkling model on Databricks offers context, control, choice, and cost advantages that directly challenge the incumbent consumption model for large language models.
Open-Weights Plus Unity AI Gateway: Control Without Friction
The real shift is not only that Inkling is an open-weights AI model; it is that Databricks has wrapped it in enterprise-grade governance through Unity AI Gateway. Inkling can be fine-tuned on proprietary codebases, internal documentation, and domain-specific data to reach higher accuracy on specific tasks, while data stays inside a governed environment. That combination of open weights and strong policy controls breaks the old trade-off where enterprises had to pick between control and convenience. Now teams can set up a governed Inkling endpoint in seconds, with security, permissions, audit logging, cost limits, and observability baked in. According to Databricks, this means no lock-in to a single model provider and the ability to switch, combine, or customize models as requirements evolve and workloads change. For cautious enterprises, this is the first time "open" does not mean "unsupported".
From API Fees to Tinker: A Business Model That Favors Developers
Inkling’s business model might matter more than its benchmark scores. The startup launched Inkling as an open-weights AI model that anyone can download, inspect, and modify for free, with weights available publicly. Instead of charging per API call like many proprietary providers, Thinking Machines will earn revenue through Tinker, a paid platform for fine-tuning Inkling to specific business tasks. That shifts spending away from opaque per-token pricing toward infrastructure and customization the enterprise controls. One analyst noted that companies will likely pay for Tinker, the platform they use to tailor Inkling for their own workloads. In effect, developers are no longer penalized for usage; they are rewarded for investing in better models tuned to their domain. This approach aligns commercial incentives with technical reality: fine-tuning is where most of the value in enterprise LLM deployment is created.
A Technical Bet on Customization, Not Raw Power
Technically, Inkling is ambitious but not obsessed with winning every leaderboard. It is a Mixture-of-Experts Transformer with 975 billion total parameters, of which only 41 billion are active for any given prompt. It was trained from scratch on 45 trillion tokens across text, images, audio, and video in less than nine months on Nvidia’s GB300 NVL72 systems. Inkling natively handles text, images, and audio with a context window of up to one million tokens and offers adjustable "thinking effort" controls to trade speed for accuracy. It even flags its own outputs for uncertainty instead of confidently hallucinating. The company openly states that Inkling is not the strongest overall model available today, and it still trails leading open models on some text and code benchmarks. But early evidence from fine-tuning work, such as financial models outperforming proprietary alternatives at a fraction of the cost, supports the thesis that customization beats chasing universal performance crowns.
What This Means for AI Agent Development and the Road Ahead
Inkling’s integration with Databricks is tailored for AI agent development, not just chatbots. Through Unity AI Gateway, teams can connect Inkling to popular coding agents like Cursor, OpenCode, or Pi and centrally govern access, budgets, and security. They can also build Inkling-powered agents with Agent Bricks to analyze data, automate complex work, evaluate results with custom judges, and deploy at scale. The Inkling model Databricks pairing lowers barriers to deploying large language models in production, offering day-zero access and removing lock-in by letting teams choose and swap models across open and proprietary options. Support to query Inkling in SQL is coming soon, which will pull more data teams into the fold. Looking ahead, Thinking Machines is previewing Inkling-Small, a 276-billion-parameter variant with 12 billion active that already beats its larger sibling on some reasoning benchmarks, with full weights to follow after testing. The direction of travel is clear: open-weights AI, integrated with governed platforms, is becoming the default canvas for serious enterprise AI systems.
