Thinking Machines Lab, the AI startup led by former OpenAI CTO Mira Murati, released its first model this week. The important part is not that Inkling is trying to win every benchmark. It is that Thinking Machines is making customization the center of the story.
Inkling is a large open-weight multimodal model: a 975B-parameter mixture-of-experts system with 41B active parameters and a context window of up to 1M tokens. Thinking Machines says it was pretrained on 45 trillion tokens across text, images, audio, and video, and that it can reason natively over text, images, and audio.
That spec sheet is large enough to matter, but the positioning matters more. Thinking Machines says directly that Inkling is not the strongest overall model available today. Instead, it is meant to be a practical open-weight base for customization: broad, multimodal, efficient, and available for fine-tuning through Tinker.
The model is the starting point
That is a useful change in emphasis. For the last two years, most model launches have been narrated as a race: bigger benchmarks, longer context, lower price, better coding score, better reasoning score. Inkling points at a different product question: what happens after the model lands inside a real workflow?
Thinking Machines is pairing Inkling with Tinker, its fine-tuning platform. The company is not only saying "use our model." It is saying: take the model, adapt it, test it, and make it behave more like the system your team actually needs.
That is why the release feels commercially important even if Inkling does not sit alone at the top of every leaderboard. Many companies do not need a general assistant that is impressive in a demo and expensive in production. They need a model that can be shaped around support policies, internal vocabulary, document patterns, code conventions, user permissions, and business-specific edge cases.
Open weights change the buying conversation
Open weights do not automatically make a model easy to run, cheap to deploy, or safe to customize. A model at this scale still demands serious engineering discipline. But open weights change the posture of the buyer and builder. They create more room for inspection, adaptation, self-hosting strategies, fallback planning, and specialized post-training.
That matters because enterprise AI buyers are becoming more sophisticated. They are asking harder questions about token bills, data control, workflow lock-in, and whether their best process knowledge should be handed permanently to a closed platform. We covered that pressure recently in the enterprise pushback against outside AI labs.
Inkling lands directly in that mood. It gives teams a different argument: do not just rent an answer engine. Build a controllable system around a base model, tune it to the work, and keep more leverage over the architecture.
Customization is becoming the product layer
The most interesting demo in Thinking Machines' announcement is not just that Inkling can answer questions or produce artifacts. It is that the company shows Inkling using Tinker to fine-tune itself toward a target behavior, then switching to the updated weights. Whether or not every team needs that exact loop, the message is clear: post-training is becoming a normal product operation.
That has consequences for app builders. A useful AI product is no longer only a prompt wrapped around a model API. It may need datasets, evals, routing, model-specific behaviors, permission controls, logs, rollback paths, and recurring improvement cycles. The model launch is not the finish line. It is the base layer.
That is also why multimodality matters here. If a model can handle text, image, and audio inputs, the customization surface gets wider. A business can tune around screenshots, transcripts, forms, product images, support calls, scanned documents, or mixed media workflows. The AI system becomes less like a chatbot and more like a configurable worker inside a bounded process.
SunMarc should read this as a workflow signal
For SunMarc App Labs, the lesson is not "use the largest model possible." The lesson is that durable AI features should be narrow, inspectable, and tuned to the job.
QR Remix does not need vague general intelligence. It needs reliable transformation logic, previewable outputs, and clear user control before a regenerated code is used. PDF Merger & Splitter does not need a talking assistant bolted onto the side. It needs document workflows that remain local, predictable, and reversible. GeoPoint-style navigation tools do not need model spectacle. They need state, intent, location use, and route decisions to be obvious.
Inkling reinforces that direction. AI value is moving toward systems that can be shaped around a specific task, cost profile, trust boundary, and user promise. In other words, smaller product truths matter more than giant demos.
The next AI race is less one-size-fits-all
The closed frontier labs will still matter. Benchmark leaders will still set expectations. But Inkling is part of a broader market shift: more customers want control, more builders want adaptable base models, and more serious AI products will be judged by how well they fit a workflow instead of how loudly they launch.
That makes Thinking Machines' first release worth watching. Inkling is not trying to end the model race. It is trying to change what the race is about: from one model for everyone to AI systems that teams can bend toward their own work.
For product builders, that is the practical takeaway. The next useful AI advantage may not be the biggest model. It may be the most adaptable system around the model.
Relevant links
- Thinking Machines: Inkling, our open-weights model
- Thinking Machines: Inkling model card
- Hugging Face: thinkingmachines/Inkling
- Axios: Mira Murati's Thinking Machines debuts first AI model
- WIRED: Thinking Machines Lab drops its first model
- SunMarc archive: Enterprise AI Buyers Are Starting to Push Back on the Labs
- SunMarc archive: The AI Copyfight Is Moving From the Web to the Model Layer