Thinking Machines Lab, the company founded by Mira Murati after leaving OpenAI, has released Inkling, its first open-weights model. It’s a Mixture-of-Experts with 975 billion total parameters and 41 billion active, trained on 45 trillion tokens (text, images, audio, video). Apache-2.0 license, available on Hugging Face and through the API of Tinker, the company’s training platform.

It’s not a frontier model: Thinking Machines says so openly. It’s a strong base model, designed for fine-tuning on Tinker. Its scale puts it in competition with open Chinese models (GLM 5.2) and open-weights USA alternatives like NVIDIA Nemotron and Gemma 4. Native context is 1 million tokens, extended to 64K/256K on Tinker.

Why it matters. A model of this scale with a permissive license is a concrete option for those who want self-hosting or fine-tuning without relying on the big players. The arrival of a new player with strong credentials (Murati was OpenAI’s CTO) and a declared focus on Asia-Pacific shifts the balance. The model card and training data documentation are sparse, though: two generic paragraphs, no details on sources.

If you want to try it. Inkling is available via API on Tinker (API key required) or downloadable from Hugging Face for self-hosting. The company also promises Inkling-Small (276B total, 12B active), but weights will only be released after testing is complete.

In detail

The context.

Mira Murati left OpenAI in 2024, after serving as CTO during the crucial period between GPT-4 and GPT-5. Thinking Machines Lab is her new company, founded with a declared focus on open models and the Asia-Pacific region. Inkling is the first release: a signal of positioning in the open-weights market, where Chinese models (GLM, Qwen, DeepSeek) have dominated in recent months.

What the documentation actually says.

The model card is brief: MoE architecture, 975B total parameters, 41B active at each inference, 45 trillion training tokens (mix of text, images, audio, video). Apache-2.0 license, so commercially usable without restrictions. 1M token native context, extended to 64K or 256K on Tinker.

The Training Data Documentation is even sparser: two paragraphs stating only that data comes from the public internet, accessible repositories, and third parties. No dataset list, no percentages per modality, no details on filters or curation. For a model positioning itself as a base for enterprise fine-tuning, this falls short of standards set by US labs (OpenAI, Anthropic, Google document much more).

The missing numbers.

There are no public benchmarks in the announcement. Thinking Machines says Inkling “is not the strongest model overall,” but provides no quantitative comparisons with Nemotron, Gemma 4, or GLM 5.2. Early independent tests of SVG generation and multimodal vision show the limits: the pelican-on-bicycle SVG is recognizable but imprecise, and the model described the pelican as “a stork or a seagull.” Interesting examples, but anecdotal.

What it means for developers.

A 975B parameter model with a permissive license is novel in the US landscape: Nemotron Ultra (3.1T) has a more restrictive license, Gemma 4 maxes out at 27B. Inkling positions itself as a self-hostable alternative to proprietary models, with manageable inference costs thanks to its MoE architecture (only 41B active). For those wanting fine-tuning on Tinker, the model is designed for it: efficient training, long context, multimodal.

The Asia-Pacific focus is explicit: Thinking Machines has servers in the region and targets a market where Chinese models already have strong presence. Competition is direct.

Declared and implicit limitations.

Thinking Machines declares that Inkling is not frontier-grade and serves as a base for customization. Technical limitations aren’t documented: no information on failure modes, known biases, or tasks where the model underperforms. The data documentation is so generic it prevents independent verification.

Inkling-Small (276B/12B) is promised but not yet released: “weights will be released once testing is complete.” There’s no timeline.

Where Thinking Machines is headed.

Inkling’s release signals a clear strategy: open models, proprietary training platform (Tinker), focus on enterprise fine-tuning. It’s the opposite business model from OpenAI (closed API, no weights) and similar to Hugging Face or Replicate. The team’s credibility (Murati, plus other unnamed ex-OpenAI members) attracts attention, but the model must be judged on numbers, not pedigree.

For now, Inkling is an interesting contender in the open-weights segment, but with substandard documentation and no benchmarks. Those evaluating it will need to test it on their own use cases, not rely on credentials.