What Is Inkling: Thinking Machines Lab’s bet on customisable AI
Follow Us
Inkling by Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has become the first open weights model from the company. The model has quite an impressive architecture – it is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion of which are active all the time. The model works with natural language, images, audio, and video formats, can work with context windows of up to 1 million tokens and was trained on 45 trillion tokens. There is also a smaller preview variant of Inkling – Inkling-Small, which uses only 12 billion active parameters.
SurveyAlso read: This OpenAI model breached a vending machine to prove a point on AI safety
And here is the part, which is more important than the specification of the model. Thinking Machines is open about the fact that Inkling is not the best available model, whether it is open or proprietary one. For example, in benchmarking tests SWE-Bench, HLE and SimpleQA it performs way below Claude, GPT and Gemini’s frontier models, and it is also not ahead of other open weights models in all of the tests. For instance, Nemotron 3 Ultra and GLM 5.2 outperform Inkling in some coding and reasoning tests.
But what’s the point in releasing it then? The key value proposition for the company is customisation rather than raw power. Inkling is available for fine-tuning on Tinker, Thinking Machines’ fine-tuning training platform, from today. The reasoning here is that a broader, well-rounded generalist model is better suited for specialisation than an optimised frontier one. Whether this holds true depends solely on how well Inkling performs under fine-tuning conditions, which cannot be assessed just through benchmarks.
This was illustrated through Inkling fine-tuning itself, creating its own fine-tuning job, generating synthetic eval data and loading back its new weights into the coding agent. It was quite a nifty demo, but a lipogram-writing trick does not say much about Inkling’s performance under a more complex real-world fine-tuning scenario.
Also read: Grok Build’s disaster shows why it trails behind ChatGPT, Claude, Gemini
There are several points worth noting. First, Inkling is capable of fine-tuning itself and controllable thinking effort, allowing developers to prioritise speed or economy over quality when running the model. Second, it has been reported to have equal performance to Nemotron 3 Ultra on Terminal Bench 2.1, using only a third of the tokens. Finally, Inkling has demonstrated good results on Cognition’s censorship-resistance eval, which Thinking Machines seems to be proud of.
Inkling can be obtained from Tinker, Hugging Face, and inference partners such as Together AI, Fireworks, Databricks, and Baseten.
For developers, the more important question is not whether Inkling is the most intelligent model to be found in the leaderboard. It is not. It is about whether having a customizable and mid-level generalist will beat the cost of using API access to the smarter model which is closed for their particular use case. This is a gamble made by Thinking Machines on behalf of the whole open weights community.
Also read: OpenAI’s first hardware device: Why not a wearable speaker tied to a necklace?
