Mira Murati's Thinking Machines Launches Open-Weight AI Model, Takes on NVIDIA Nemotron

Thinking Machines Lab, the AI startup founded by former OpenAI Chief Technology Officer Mira Murati, has released Inkling, its first open-weight foundation model.

Inkling is a mixture-of-experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters. The multimodal model supports a context window of up to one million tokens and was pre-trained on 45 trillion tokens spanning text, images, audio and video. 

Alongside the flagship model, the company also introduced Inkling-Small, a preview version with 276 billion total parameters and 12 billion active parameters.

“Our mission is to build AI that extends human will and judgment,” the company said in a blog post. “Today we are advancing our mission by releasing a model we trained from scratch with the full weights available, so that people can make it their own.”

Inkling is available on Thinking Machines’ Tinker platform with 64,000-token and 2,56,000-token context options for fine-tuning. The company has also launched an Inkling Playground for developers and is offering the model at a 50% introductory discount for a limited period. 

Full model weights have been released on Hugging Face, including an NVFP4 checkpoint optimised for NVIDIA Blackwell systems. The model is also available through Together AI, Fireworks, Modal, Databricks and Baseten, with support for inference frameworks including SGLang, vLLM, TokenSpeed, llama.cpp and Hugging Face Transformers.

Thinking Machines positioned Inkling as a foundation model for customisation rather than a model optimised solely for benchmark leadership.

“Inkling is not the strongest overall model available today, open or closed,” the company said. “Instead, a combination of qualities makes it a good open-weights base for customisation: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.”

Thinking Machines Lab said Inkling is now available for fine-tuning on its Tinker platform to make model customisation accessible across a wider range of use cases.

The model natively supports reasoning across text, images and audio while allowing developers to adjust its reasoning effort to trade off performance, latency and inference cost. 

According to the company, Inkling achieves performance comparable to NVIDIA’s Nemotron 3 Ultra on Terminal Bench 2.1 while generating roughly one-third as many tokens. It trails GLM 5.2 and Kimi K2.6 on the benchmark.

Thinking Machines said Inkling was trained on NVIDIA GB300 NVL72 systems and relied on more than 30 million reinforcement learning rollouts during post-training. 

The company said reasoning performance improved steadily throughout the reinforcement learning process while the model also learned to produce more concise reasoning traces without affecting final answers.

The company said Inkling was built as a general-purpose model covering reasoning, coding, agentic tool use, instruction following, factuality, vision and audio instead of specialising in a single domain. 

It also said the model ranks among the stronger open-weight models on agentic coding benchmarks and Design Arena’s Agentic Web Development leaderboard.

For multimodal tasks, Inkling accepts image and audio inputs without dedicated encoders. The model can transcribe speech, answer questions about recordings, analyse charts and diagrams, and use Python tools for image understanding, including zooming and cropping.

Thinking Machines said it trained Inkling to improve factual calibration, instruction following and resistance to censorship. The company also said external safety testing found the model performed well on benchmarks measuring refusal of harmful requests while avoiding unnecessary refusals of benign prompts.