Thinking Machines Lab Releases Inkling, an Open 975B-Parameter Model

Thinking Machines Lab, a startup led by a former technical director of OpenAI, has released the weights for Inkling, a 975-billion-parameter multimodal model. Rather than targeting top scores against closed proprietary systems, the company pitched the release as a customizable foundation for organizations that prefer running internal infrastructure over paying per API token.

The system uses a Mixture-of-Experts architecture modeled after DeepSeek-V3. Each layer contains 256 routed experts alongside two shared ones, activating six experts per token to yield 41 billion active parameters during inference. The model was trained on 45 trillion tokens spanning text, images, audio, and video, and supports a context window of up to 1 million tokens. Instead of relying on third-party encoders, Inkling converts audio directly into dMel spectrograms and visual input into 40x40 pixel patches via a four-layer hMLP, embedding them alongside text tokens.

According to benchmark data published by the lab, the model scored 43.9% on SimpleQA Verified—surpassing GLM 5.2 (38.1%) and Nemotron 3 Ultra (32.4%)—and reached 97.1% on AIME 2026. On the Terminal Bench 2.1 coding benchmark, the developers claim an adjustable reasoning parameter allowed the system to match Nemotron while consuming roughly one-third of the tokens. Early bootstrap data reportedly included synthetic output from Moonshot AI's Kimi K2.5.

During reinforcement learning over more than 30 million steps, the authors reported an unexpected efficiency artifact: the model's internal reasoning tokens condensed into an abbreviated, telegraphic style without degrading task accuracy.

While the weights are freely downloadable on Hugging Face, actually using the flagship release requires enterprise-grade hardware. The full BF16 checkpoint demands roughly 2 TB of VRAM across a GPU cluster. A quantized NVFP4 release designed for NVIDIA Blackwell architecture lowers the footprint, but still requires approximately 600 GB of VRAM. A smaller preview variant, Inkling-Small, features 12 billion active parameters out of 276 billion total.

Source: habr

Primary source: Thinking Machines Lab