Inkling (xhigh)

Inkling (xhigh) is a high-reasoning configuration of the flagship multimodal model developed by Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. Released in July 2026, Inkling is a massive 975 billion parameter sparse Mixture-of-Experts (MoE) system that natively processes text, image, and audio inputs. The "xhigh" designation represents the model's maximum reasoning effort setting, which utilizes a "controllable thinking" mechanism to allocate additional computation for complex problem-solving in domains like software engineering and mathematics.

The model's architecture consists of a 66-layer decoder-only transformer with 256 routed experts and 2 shared experts per MoE layer. During inference, only 41 billion parameters are active per token, striking a balance between the intelligence of a trillion-parameter scale model and the throughput of a much smaller one. Inkling is notably encoder-free, integrating visual and auditory signals directly into the transformer blocks using hMLP-based patch embeddings for images and dMel spectrogram projections for audio. It supports a context window of up to 1,000,000 tokens, enabling the processing of extensive document sets and long-form audio recordings.

A core focus of the Inkling project is "epistemics"—the model's ability to remain well-calibrated and resistant to censorship. It is designed to express nuanced confidence levels in its outputs and provide direct, factual information on sensitive or controversial topics that are often restricted in other frontier models. This makes it a significant open-weights alternative for enterprises requiring high-fidelity, uncensored data processing and domain-specific fine-tuning.

Inkling excels in agentic workloads and tool use, featuring built-in multi-token prediction (MTP) layers that act as internal drafters for speculative decoding. This architectural choice significantly increases generation speed without sacrificing accuracy. The model is released under an Apache 2.0 license, and Thinking Machines provides a dedicated platform called Tinker for developers to fine-tune and adapt the model for specialized private applications.