Thinking Machines offers up to $50,000 in credits for Inkling safety research
Thinking Machines Lab will subsidize outside safety work on Inkling, although it has not disclosed the program's budget or number of awards.
By RuntimeWire Staff · Published
Primary source: Aligned News - AI Intelligence
Why it matters
Murati is funding outsiders to test the central risk in Thinking Machines' customization strategy: whether users can fine-tune away safeguards faster than researchers can make them persist.
Mira Murati's Thinking Machines Lab is offering up to $50,000 in Tinker credits for outside research into the safety of its open-weight Inkling models.
The August 24 grant announcement, summarized the following day by Aligned News, follows the release of Thinking Machines' first two open-weight models. The 975-billion-parameter Inkling arrived on July 15, and the smaller 276-billion-parameter Inkling-Small followed on July 30.
Inkling is a mixture-of-experts model with 41 billion active parameters and a context window of up to 1 million tokens. Inkling-Small has 12 billion active parameters and the same maximum context length. Both can process text, images and audio, and both are available for fine-tuning through Tinker.
The program's announcement lists work on defensive capability, tamper-resistant safeguards, alignment failures, reward hacking, and measuring how risk changes after fine-tuning. Thinking Machines also invites proposals outside those categories if additional Tinker credits would accelerate the work.
Murati is placing outside researchers inside the central tension of her product strategy. Thinking Machines wants developers to own and customize capable models, a thesis built into Tinker, its managed training API. Releasing weights also lets developers remove behavioral safeguards or train models on hazardous material. The grant program pays researchers in access to the same fine-tuning infrastructure that makes those modifications easier.
Thinking Machines is asking researchers to use fine-tuning as a controlled stress test: train models toward defensive cyber skills, test whether safeguards survive adversarial modification, measure when narrow training causes broader misalignment, and forecast how risk changes as training budgets increase.
Murati is building around customization
Murati founded Thinking Machines in February 2025 with a group that included John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch and Luke Metz, according to Contrary's company profile. She had served as OpenAI's chief technology officer and left the company in September 2024 before starting the new lab.
Her argument since starting Thinking Machines has been consistent: a small number of labs should not permanently decide how one fixed AI system behaves for everyone. In a July essay, Thinking Machines said organizations should be able to shape models with their own knowledge and keep adapting them as that knowledge changes.
Tinker is the commercial and technical expression of that view. Researchers control training data, loss functions and training loops, while Thinking Machines handles distributed infrastructure, scheduling and recovery. Tinker uses LoRA, which trains smaller adapters instead of updating every parameter in the base model, allowing the platform to share computing capacity across jobs.
The safety program extends that customization thesis. Thinking Machines argued in its open-weight safety framework that downloadable models can widen access while creating irreversible misuse risks. The grants widen the pool of researchers able to test that conclusion.
The $50,000 is platform credit
The maximum award is up to $50,000 in Tinker credits per participant, measured at Tinker's current rates and usable only for the approved research proposal. It is not unrestricted cash. Unused credits expire 12 months after provisioning, according to the program terms.
Thinking Machines says it will work with selected participants for up to six months, providing model access, technical support and scheduled check-ins. Applicants must submit a one-page project summary, identify principal investigators and contributors, attach CVs, and provide organizational and tax information where applicable. The application page sets a September 25 deadline at 11:59 p.m. Pacific time and says proposals will be reviewed within one week after the deadline.
Selection will consider a proposal's relevance, feasibility, construct validity, and simplicity or generalizability. The cited application and terms pages do not disclose how many participants will receive credits or the program's total budget.
The eligibility rules require applicants to be at least 18 and exclude certain government employees, Thinking Machines personnel and their families, sanctioned or restricted people, and residents of embargoed or designated countries. The terms also address applications submitted for companies, academic institutions, research institutes and other entities, without presenting those categories as an exhaustive list of eligible applicants.
Thinking Machines claims no ownership of papers, datasets or other work product created by participants. Participants grant Thinking Machines a perpetual, irrevocable, worldwide, transferable, sublicensable, nonexclusive and royalty-free license to use that work, including for training and improving products. Published or disclosed work must use Creative Commons Attribution 4.0 or another company-approved open-source license. Work that references or uses Thinking Machines materials requires written approval before publication, according to the program terms.
The arrangement gives researchers access to managed training infrastructure that can be costly to reproduce independently. Thinking Machines receives research findings and broad rights to use the resulting work, while the experiments run on Tinker.
Grants are becoming a Tinker distribution channel
This is Thinking Machines' third grant initiative tied to the platform. In October 2025, it began offering research grants starting at $5,000 and teaching credits of $250 per student. Early recipients included classes at Stanford and Carnegie Mellon and a Stanford lab fine-tuning small-molecule chemistry models. In May 2026, Thinking Machines offered multiple interactivity research grants worth $100,000 in funding plus $25,000 in Tinker credits.
The pattern makes grants a distribution channel for Tinker alongside its research purpose. Each program recruits specialists to build training workloads, exposes the API to demanding users and produces examples for other researchers evaluating the platform. The safety program applies that approach to the hardest objection facing Murati's push for customizable AI: users can customize away the protections.
Thinking Machines has ample capital to fund the work. Murati raised a $2 billion seed round in July 2025 led by Andreessen Horowitz, with Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street participating, at a reported $12 billion valuation, according to TechCrunch.
The consequential undisclosed figure is the number of awards. A small group would produce a limited set of case studies. A larger cohort could give Thinking Machines an external testing network for Inkling and Tinker, with evidence about where open-weight releases break and which defenses survive determined fine-tuning.