975B Parameters, Full Weights, Free to Fine-Tune — Thinking Machines Just Changed the Open Model Race
Thinking Machines released Inkling with 975B parameters, full weights, and free-to-fine-tune access. This multimodal model offers a 1M-token context window for builders.
975B Parameters, Full Weights, Free to Fine-Tune: A New Open Standard
Thinking Machines just changed the open model race with a bold release. The new Inkling model features 975B parameters, full weights, and is free to fine-tune immediately. Unlike restricted releases, this Mixture-of-Experts model allows commercial use from day one via the Tinker platform.
Inkling activates 41B parameters per token and was trained on 45 trillion tokens. It handles text, images, audio, and video natively. Therefore, it serves as a practical starting point for developers who want true ownership of their AI infrastructure.
Architecture and Performance Details
Inkling uses a Mixture-of-Experts Transformer architecture similar to DeepSeek-V3. Each layer contains 256 routed experts and two shared experts. However, only six routed experts activate per forward pass to maintain efficiency.
The model supports a massive 1 million token context window. Training utilized a hybrid Muon and Adam optimizer strategy for better convergence. Consequently, Inkling ranks competitively on agentic coding and reasoning benchmarks while handling native multimodal inference without adapters.
Controllable Effort for Cost Efficiency
Benchmarks grab headlines, but production costs matter more for teams. Inkling includes a tunable effort parameter to control reasoning depth. At low effort settings, inference is fast and cheap.
At higher settings, it matches top open-weight models. Tests show it reaches high performance levels with roughly one-third the token count. This efficiency benefits developers running iterative agentic loops or local inference on constrained hardware.
Native Multimodal Capabilities
Inkling processes audio and images natively rather than through bolted-on adapters. Audio enters as spectrograms while images process as pixel patches. Both modalities embed jointly with text tokens from the start.
This design powers real-time interaction systems for voice and vision. Benchmark results confirm strong performance on specialized audio and chart reasoning tasks. Thus, the model functions effectively as a background reasoning engine for complex user interactions.
Safety and Calibration Focus
Thinking Machines prioritized calibration and censorship resistance during training. Reinforcement learning helps the model admit uncertainty instead of confabulating answers. This results in forecasting scores comparable to leading proprietary models.
Safety evaluations show Inkling leads the open-weight field on adversarial tests. It refuses harmful requests reliably while avoiding unnecessary refusals on benign topics. First, this balance ensures the model remains useful and safe for diverse applications.
Fine-Tuning and Future Roadmap
The true value lies in the surrounding Tinker platform. Developers can test base behavior before committing to training runs. Thinking Machines even demonstrated Inkling fine-tuning itself within an agentic harness.
This release is just the beginning of a new model family. An upcoming Inkling-Small variant targets lower-latency use cases. While not the absolute strongest model available, its openness makes it uniquely buildable compared to competitors like Qwen3.8’s 2.4T Parameters Are Already Live — and Open Weights Are Coming.