Thinking Machines Lab releases Inkling, an open-weight AI model with 975 billion parameters
The San Francisco-based startup's Inkling model is open-weight and built partly using Chinese AI techniques, targeting enterprise customization over raw benchmark performance.
What Inkling is and why it matters
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first general-purpose model on Wednesday. Named Inkling, the 975-billion-parameter system is open-weight, meaning developers can download, run and modify it. The release gives Western enterprises a new alternative to Chinese open-source models, which have dominated that segment since Meta abandoned its open Llama strategy. The model is available on the company's Tinker fine-tuning platform and on Hugging Face.
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.
Technical details and Chinese influence
The model was built as a mixture-of-experts architecture, drawing on roughly 41 billion of its 975 billion total parameters for a given task. It was trained on 45 trillion tokens of text, image, audio and video. The architecture borrows from Chinese lab DeepSeek-V3, according to the Financial Times, and its final post-training phase used data generated by Moonshot AI's Kimi K2.5, a Beijing-based model. Researchers also discovered during training that Inkling began dropping natural language reasoning steps to save compute, prompting the team to reinstate them for explainability.
It determined that the grammar was overhead, which is interesting.
Performance vs. the competition
Thinking Machines is upfront that Inkling is not the strongest model available. Benchmarks shared with reporters place it below closed models from Anthropic and OpenAI and behind several top Chinese open models. However, the company highlights competitive results on agent-related tasks and notes that Inkling uses one-third as many tokens as Nvidia's Nemotron 3 Ultra to hit equivalent coding performance. The model also allows users to dial "thinking effort" up or down, trading speed for depth.
The startup behind it
Mira Murati co-founded Thinking Machines in February 2025 after leaving OpenAI, where she had briefly served as interim CEO. She raised a $2 billion seed round at a $12 billion post-money valuation from Andreessen Horowitz, Nvidia, AMD and hedge fund Jane Street. The company launched its Tinker customization platform in October 2025 and has been generating revenue through enterprise fine-tuning deals. Earlier this year it suffered a talent drain when several senior staffers departed for Meta and OpenAI, questioning Murati's leadership.
- Mira Murati and former OpenAI colleagues found Thinking Machines Lab
- Company closes a $2 billion seed round at a $12 billion valuation
- Tinker model-customization platform launches for enterprise clients
- Research preview of interaction models that listen and speak natively
- First general-purpose model, Inkling, released as open-weight
Open-weight strategy and market context
Demand for open-weight models is growing as enterprises seek cheaper, self-hosted AI they can tailor. Palantir CEO Alex Karp recently argued on CNBC that frontier closed models are too expensive and lack clear IP protections. Thinking Machines' bet is that customizability, not maximum intelligence, will win enterprise accounts. The company said it is already training more powerful successors and is previewing a lighter-weight variant called Inkling-Small.
