Thinking Machines Lab, founded in early 2025 by former OpenAI CTO Murati, released its first AI model on Wednesday, July 152. Code-named Inkling, the open-weights model weighs in at 975 billion parameters, making it the largest American open-weights model to date.
Inkling requires more than two terabytes of GPU memory to run at its native 16-bit precision — a quantity present in around eight of Nvidia's B300 accelerators, or sixteen H200s. Thinking Machines has also released an NVFP4 quantized version of the model capable of running on half the GPUs.
The model was trained to understand video and audio3. Thinking Machines says Inkling is competitive with Chinese open models such as DeepSeek V4, GLM 5.2, and Kimi K2.6 in terms of size and capabilities across a variety of workloads, though its benchmark charts also show it trailing proprietary models like Anthropic's Claude and OpenAI's GPT. The Register's Tobias Mann cautioned readers to "take these claims with a grain of salt — gaming AI benchmarks isn't exactly difficult".
Together AI announced it is hosting Inkling on day zero of the model's release4.
ANALYSIS The release positions Thinking Machines as a direct competitor to the Chinese labs that have dominated the large-scale open-weights tier, while also differentiating from U.S. frontier labs like OpenAI and Anthropic, which have kept their most capable models proprietary. The availability of a quantized variant lowers the hardware floor, potentially broadening the model's adoption among organizations that lack access to large GPU clusters.
The multimodal training scope — video and audio in addition to text — signals that Thinking Machines is not pursuing a text-only strategy, which could matter for downstream applications in media analysis and robotics.
Inkling's official announcement was posted on Thinking Machines' website5,6. The Hacker News listing had accumulated 84 points and 15 comments on one thread, and 51 points and 8 comments on another, as of the time of capture.