Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released its first model today: Inkling, a 975-billion-parameter open-weights multimodal system that anyone can download, run, and fine-tune. The launch, confirmed hours ago by the company alongside reporting from Bloomberg, Axios, and TechCrunch, marks Murati's public debut as a model builder and lands a rare Western entry into an open-weights field that Chinese labs had come to dominate.
- Inkling is a sparse Mixture-of-Experts: 975B total parameters but only about 41B active per token, spread across 256 experts, so it runs far cheaper than its raw size suggests.
- It is multimodal and long-context: a 1M-token window, pretrained on 45 trillion tokens of text, images, audio, and video, able to reason over three modalities at once.
- The weights are actually open: the full model is live on Hugging Face and fine-tunable now on Tinker, the company's customization platform, with a lighter Inkling-Small (12B active) preview to follow.
- Thinking Machines is not claiming the crown: it says Inkling is not the most powerful model available, closed or open, but is competitive across agentic tasks with no glaring weak spots.
What did Thinking Machines actually release?
Inkling is a decoder-only, multimodal Mixture-of-Experts model. The headline number is 975 billion total parameters, but the design is deliberately sparse: for any single token, the router activates only about 41 billion parameters out of 256 experts. That is the same trick DeepSeek, Kimi, and other large open models use to stay fast and affordable, and it is why a near-trillion-parameter model can serve at a fraction of the cost a dense model that size would demand.
RelatedKAT-Coder-Pro V2.5 Lands Second on SWE-Bench Pro
The model was pretrained on 45 trillion tokens spanning text, images, audio, and video, and it carries a 1M-token context window. Thinking Machines says it can reason over all three input modalities simultaneously rather than bolting a vision encoder onto a text core. Crucially, the full weights are published on Hugging Face and the model is already live for fine-tuning on Tinker, the company's own customization API. Inkling is billed as the first in a family: a smaller preview, Inkling-Small, uses 12B active parameters and will get its weights after further testing.
Why does an open-weights model from Murati matter?
The strategic story is geography. Reuters framed Inkling explicitly as a Western alternative to the widely used open models coming out of Chinese labs. That gap became real over the past year: after Meta's lackluster Llama 4 release, the company pulled back from its open Llama line, and many teams that wanted downloadable, self-hostable weights ended up defaulting to Chinese options like Qwen, DeepSeek, and Kimi. Inkling is the most credible US-based answer to arrive since.
The appeal of open weights is concrete. Companies can fine-tune the model on proprietary data, run it on infrastructure they control, and avoid per-token API costs and vendor lock-in. Thinking Machines leaned into that with a live demo of self-fine-tuning: it gave Inkling a target behavior that prompting alone cannot reliably enforce (a lipogram model that never uses the letter "e"), and the model drafted the plan, generated its own evaluation and synthetic data, called the Tinker API to post-train itself, loaded the new weights, and closed the loop as a freshly customized model.
| Model | Inkling | Typical Chinese open MoE | Meta Llama (legacy) |
|---|---|---|---|
| Origin | Thinking Machines (US) | Qwen, DeepSeek, Kimi | Meta (pulled back) |
| Weights | Open, on Hugging Face | Open | Open, aging |
| Design | 975B total / 41B active MoE | Sparse MoE | Mostly dense |
| Modalities | Text, image, audio, video | Varies, often text-first | Text, some vision |
| Context | 1M tokens | 128K to 1M | 128K |
| Fine-tune path | Tinker API, self-tuning demo | Community tooling | Community tooling |
How good is it on benchmarks?
Here Thinking Machines is unusually candid. It states plainly that Inkling is not the most powerful model currently available, closed or open. Instead of chasing a single leaderboard peak, the company trained it broadly across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, arguing that a well-rounded base is what actually matters for customization. Reuters echoed that read: Inkling does not lead the pack overall but performs competitively, especially on agent tasks where a model plans and takes multi-step actions.
The metric the company chose to highlight is efficiency, not raw scores. It published an effort-versus-performance curve on Terminal Bench 2.1 (agentic coding), HLE (advanced reasoning), and IFBench (instruction following), with controllable "thinking effort." Its sharpest claim: Inkling spends roughly one-third as many tokens to reach the same Terminal Bench performance as Nemotron 3 Ultra. For anyone paying per token in production, that is the number that matters more than a benchmark bragging point.
RelatedCursor Builds 'Sand' Agent to Rival Claude Cowork
What does it mean for the market?
Thinking Machines is private and reportedly valued in the tens of billions on the strength of Murati's team, so there is no ticker to trade here. The signal for investors sits upstream and downstream. Upstream, a credible Western open-weights model pressures the API pricing of closed vendors: if enterprises can self-host a capable multimodal model, some of the volume that would have flowed to OpenAI, Anthropic, and Google gets recaptured on their own hardware. Downstream, it is bullish for the fine-tuning and inference tooling layer, exactly where Tinker plays, and for the compute providers that host open models. The read is not "buy anything," it is that the open-weights tier just gained a serious US anchor, and that reshapes where AI budgets can go.
Who is affected first?
Enterprise teams that wanted downloadable weights but balked at deploying a Chinese model for compliance or procurement reasons now have a Western option they can audit and control. Fine-tuning shops and platforms benefit from a fresh, permissively available base. And the broader open ecosystem, which had thinned since Meta's retreat, gets a shot of momentum from a founder whose pedigree guarantees attention.
- Independent benchmarks. Vendor curves are a starting point; watch third-party evals on SWE-bench, GPQA, and agent suites to see where Inkling really lands versus Qwen and DeepSeek.
- Inkling-Small weights. The 12B-active preview is the one most teams will actually run; its release and license terms will decide real-world adoption.
- License fine print. "Open weights" covers a wide range; the exact commercial terms determine whether startups can build on it freely.
- Tinker traction. The self-fine-tuning demo is the differentiator; whether customization on Tinker becomes a habit is the business question.
Our take
Inkling is a smart first move, not a knockout blow, and Thinking Machines seems to know it. By refusing to claim state-of-the-art and instead selling breadth, controllability, and token efficiency, the company is aiming at the part of the market that closed frontier labs underserve: teams that need to own and shape a capable model rather than rent the very best one. Whether it displaces the Chinese open models depends on the license and on Inkling-Small landing well. But as a statement that the West still builds open models, and as Murati's opening argument, it is a strong one.
- OfficialThinking Machines Lab — Inkling: Our Open-Weights Model the announcement, architecture, and efficiency curves
- WeightsHugging Face — Welcome Inkling by Thinking Machines model card and downloadable weights
- ReportingTechCrunch — Thinking Machines' first open model, Inkling positioning and customization angle
- ReportingAxios — Mira Murati's Thinking Machines debuts its first AI model launch confirmation and context
Original analysis by GenZTech. Details current as of July 2026. Reporting via Reuters.