Thinking Machines releases Inkling, 975B-parameter open model

Thinking Machines ships first foundation model with open weights

Thinking Machines has released Inkling, a 975-billion-parameter mixture-of-experts transformer model with full weights available for developers to customize. The model features 41 billion active parameters, supports context windows up to one million tokens, and was pretrained on 45 trillion tokens spanning text, images, audio, and video, according to an announcement from the company.

The release includes a preview of Inkling-Small, a lighter 276-billion-parameter variant with 12 billion active parameters that matches or exceeds its larger counterpart on several benchmarks while offering lower cost and latency.

Why it matters

Thinking Machines is positioning Inkling not as the strongest overall model available, but as an optimized foundation for customization. This strategy targets a gap in the market: organizations need models they can adapt to specific workflows rather than general-purpose systems optimized for public benchmarks. The company's emphasis on controllable thinking effort—allowing developers to balance performance against token cost—addresses a practical constraint that becomes critical when running models millions of times in production environments.

Controllable compute and multimodal capabilities

Inkling implements what Thinking Machines calls "controllable thinking effort," enabling developers to adjust how many tokens the model uses for reasoning. On Terminal Bench 2.1, an agentic coding benchmark, Inkling uses one-third as many tokens as Nemotron 3 Ultra to achieve equivalent performance, the company reported.

The model processes text, images, and audio natively through an encoder-free architecture. Audio signals are input as dMel spectrograms, while images are encoded as 40x40 pixel patches. Thinking Machines designed these multimodal components to support its interaction models system, which enables real-time collaboration using voice and vision.

Training approach and safety measures

The model was trained using a hybrid optimization strategy combining Muon for large matrix weights and Adam for other parameters. Post-training involved large-scale reinforcement learning scaled to over 30 million rollouts, with reasoning performance improving log-linearly throughout the process.

Thinking Machines trained Inkling against what it calls "censorship non-compliance," aiming for direct answers on potentially sensitive topics. External testing by Cognition found strong patterns of censorship resistance. On safety benchmarks, the model scored above 98% on StrongREJECT, which tests refusal of unambiguous harmful requests, while showing the strongest built-in safeguards among compared open-weights models on FORTRESS, a benchmark testing weapons and violence-related request refusals.

Availability and fine-tuning

Both Inkling and Inkling-Small are available for fine-tuning on Tinker, Thinking Machines' customization platform. The company added an Inkling Playground in the Tinker console for developers to interact with the model before committing to fine-tuning workflows.

The models were trained on NVIDIA GB300 NVL72 systems, with Thinking Machines indicating future releases will push compute scale further across pretraining, post-training, and reinforcement learning phases.

Details were first reported by Thinking Machines in a company announcement.