Inkling Redefines Enterprise AI Customization From Day One
Inkling is an open-weights AI model designed as a downloadable, customizable multimodal Mixture-of-Experts transformer that enterprises can fine-tune, govern, and deploy on their own infrastructure instead of renting a generic hosted system. That framing matters: this is not a demo chatbot, but a direct challenge to the assumption that serious AI must live behind someone else’s API. Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released Inkling as its first in-house model and made it available for anyone to download and modify. Databricks is a day-zero launch partner, bringing the Inkling model to its platform so customers can apply it to enterprise data and coding workflows without losing control to a single vendor. The headline shift is simple: weights, not endpoints, are the product.
Inside Inkling’s Architecture: Scale, Multimodality and Open Weights
On paper, Inkling is unapologetically ambitious. It is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token, a 1 million token context window, and pretraining on 45 trillion tokens spanning text, images, audio and video. Full weights sit on Hugging Face, including an NVFP4 checkpoint for efficient inference on Blackwell-class systems, which means the community can pull, inspect and customize the exact model that Thinking Machines trained. Multimodality is not bolted on; audio is fed as dMel spectrograms and images as 40×40 patches into a lightweight MLP, then processed jointly with text tokens. The current outputs are text-only, but the architecture is clearly built for multimodal AI deployment. A smaller sibling, Inkling-Small, runs 276 billion parameters with 12 billion active and already matches or beats the main model on some benchmarks, though its weights are still in testing and will come later.
Benchmarks and Calibration: Good Enough to Be Worth Owning
Thinking Machines is candid that Inkling is "not the strongest overall model available today, open or closed," but the numbers show it belongs in the top open-source tier. Chat performance hits 79.8% on IFBench, ahead of every comparison model listed, including Claude Fable 5 and GPT 5.6 Sol. Coding and agentic work land at 77.6% on SWE-bench Verified, 63.8% on Terminal Bench 2.1, and 74.1% on MCP Atlas. Multimodal scores include 56.6% on Audio MC, 77.2% on MMAU and 73.5% on MMMU Pro, with vision reaching 82.0% when the model can call Python tools. The more interesting bet is calibration: reinforcement learning on resolved questions yields a 61.1 Brier score on ForecastBench without search, level with leading proprietary systems, while safety scores hit 78.0% on adversarial FORTRESS and 98.6% on StrongREJECT. Inkling’s strength is not perfection; it is being competitive enough that owning it becomes rational.
Open Weights, Not Open Charity: Tinker and Databricks as the Business Model
Inkling’s open-weights release is not philanthropy; it is a distribution strategy. Once full weights are public, no one is forced to pay for inference, so revenue must come from the Tinker fine-tuning platform and from hosting partners. Tinker offers Inkling with 64K and 256K context options at a 50% discount for a limited time, plus an Inkling Playground for experimentation before serious runs. The company positions Inkling not as a finished product but as a starting point for custom models that enterprises shape themselves. This bet is not theory. In a project with Bridgewater Associates, a fine-tuned open model built on specialized financial data outperformed proprietary systems on leading financial reasoning benchmarks at under 10% of their cost. According to the company, "there is evidence the bet has legs," and the evidence here is a real-world deployment, not a marketing deck.
Inkling on Databricks: Vendor-Neutral AI as a Strategic Weapon
The Inkling model Databricks launch turns open weights into something enterprises can use immediately. Inkling is available through the Unity AI Gateway, which brings shared security, permissions, audit logging and policy enforcement, so data and governance stay inside existing controls. Open weight models like Inkling can be fine-tuned on proprietary codebases, internal documentation and domain-specific data to hit higher accuracy on specific tasks, while avoiding per-token API pricing and letting teams scale deployments as needed. In other words, this is enterprise AI customization without vendor lock-in: teams can switch, combine or replace models as requirements evolve, and still plug them into the same platform. SQL querying support for Inkling is on the roadmap, widening access for analysts. If there is a single strategic takeaway, it is this: for enterprises that treat AI as infrastructure, owning a capable open-weights AI model plus a fine-tuning platform is starting to look more defensible than renting a one-size-fits-all API.
