Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released its first public model on July 15, called Inkling. Unlike the flagship systems from OpenAI, Anthropic, or Google, Inkling ships with open weights, meaning outside developers can download, inspect, and modify it directly rather than accessing it only through an API.
Inkling is a 66-layer, decoder-only Mixture-of-Experts model with 975 billion total parameters, though only around 41 billion are active for any given token, a design that keeps very large models faster and cheaper to run than their full size would suggest. Each Mixture-of-Experts layer contains 256 routed experts alongside two shared experts. The model was trained from scratch on roughly 45 trillion tokens spanning text, images, audio, and video, and it natively reasons across all four input modalities, though its outputs for now are limited to text, including code, structured data, and styled content. It supports a context window of up to one million tokens and offers controllable thinking effort, letting users trade response speed for depth of reasoning depending on the task.
Thinking Machines is notably candid about where Inkling stands competitively. The company's own release materials state plainly that Inkling is not the strongest overall model available today, open or closed. Independent evaluation from Artificial Analysis's Intelligence Index nonetheless ranked it as the strongest open-weights model released by a US lab at the time of launch. Thinking Machines frames its value differently: as a broad, balanced foundation model designed specifically to be customized, rather than one built to top every leaderboard. On one internal benchmark, the company says Inkling uses roughly a third as many tokens as Nvidia's Nemotron 3 Ultra to reach comparable coding performance.
The model ships with day-zero support in the transformers library and major inference engines including vLLM, SGLang, and llama.cpp. Given its size, running it is not casual: the full bf16 checkpoint requires roughly 2 terabytes of VRAM, while a well-calibrated NVFP4 quantized variant brings that down to around 600 gigabytes, still a substantial multi-node deployment for most organizations. Inkling is also available through Tinker, Thinking Machines's fine-tuning service, with context length options of 64K and 256K tokens for that workflow, alongside third-party inference providers.
A smaller sibling, Inkling-Small, with 12 billion active parameters and the same controllable thinking effort, is described as a natural fit for latency- and cost-sensitive workloads such as coding assistance, using an LLM to grade other models' outputs, or generating synthetic training data. Its full weights had not yet been released at launch, with Thinking Machines saying testing was still being finished.
On safety, Thinking Machines says it evaluated Inkling for biological, chemical, and cyber uplift risk using internal evaluations, external testing, and refusal-suppressed variants meant to estimate latent capability with safeguards removed, concluding the model does not present risk of material uplift beyond what is already available in the broader open-weight ecosystem. The company notes the model shows some tendency to comply with role-play or indirectly framed prompts around harmful topics, a pattern it says is consistent with other open-weight models and best addressed with layered moderation tools like Llama Guard rather than relying solely on the model's own refusals.