Inkling, Thinking Machines' first open-weights model, favors custom solutions over the podium

Thinking Machines delivers Inkling, a 975-billion-parameter MoE multimodal model. An optimized base for custom training.

Thinking Machines Lab is releasing the full weights of its first model, Inkling, trained from scratch. It is a 975-billion-parameter Mixture-of-Experts transformer, with 41 billion active parameters and a context window of up to 1 million tokens, pre-trained on 45 trillion tokens of text, images, audio, and video. A smaller preview sibling, Inkling-Small, reduces the active portion to 12 billion parameters for lower cost and latency.

The lab is taking a counter-current positioning: Inkling is not presented as the most powerful model currently available, open or closed, but as a solid foundation to customize. What makes it a good base, according to the company, is its native multimodality (it reasons across text, images, and audio), an adjustable thinking effort that balances cost and performance, and its availability for fine-tuning on Tinker. To illustrate this, the team asked the model to fine-tune itself: Inkling wrote its own training job, launched it, and evaluated the result.

On the design side, the MoE architecture closely follows that of DeepSeek-V3 and swaps RoPE for a relative positional embedding, deemed better for long sequences. Training utilized NVIDIA GB300 NVL72 systems and a RL phase scaled to over 30 million rollouts, during which the chain of thought spontaneously compressed, losing articles and connectors while remaining understandable—an effect the reward did not target. The model is accessible now on Tinker, at half price for a limited time, with its weights uploaded to Hugging Face and served via Together, Fireworks, Modal, Databricks, and Baseten.