Thinking Machines Launches Inkling, Its First

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released its first AI model. It's called Inkling, and the company would like you to know upfront that it is not the most powerful model available. That's an unusual opening pitch for a lab sitting on a reported $2 billion in funding at a $12 billion valuation—but it's also, as it turns out, the honest one.

Inkling is a mixture-of-experts (MoE) transformer released as a fully open-weight model under the Apache 2.0 license. According to Thinking Machines' own announcement, it has 975 billion total parameters, though because of the MoE architecture only around 41 billion activate for any given prompt—keeping inference costs manageable. It supports a context window of up to one million tokens and was pre-trained across text, images, audio, and video, making it natively multimodal at the reasoning level, not just at the input layer. The weights are available on Hugging Face now.

Per the AI Revolution video covering the launch, the whole thing went from zero to release in under nine months—fast by any measure, genuinely remarkable when you consider that OpenAI and Anthropic operate on multi-year development cycles. Thinking Machines' only prior product was Tinker, a fine-tuning API that launched last October. Inkling is the base model Tinker was always waiting for.

What It Can and Can't Do

On headline benchmarks, Inkling trails the field. The video cites scores from Thinking Machines' own release materials: on Humanity's Last Exam (text only), Inkling scores 29.7%, compared to significantly higher scores from competing models. On SWE-Bench Verified it lands at 77.6%—respectable, but not leading. The company doesn't dispute any of this. What it argues instead is that Inkling was deliberately trained as a generalist: balanced across reasoning, coding, instruction-following, vision, audio, and agentic tasks, rather than tuned to dominate any single leaderboard.

The efficiency case is more interesting than the raw scores suggest. According to the video's reading of Thinking Machines' release data, Inkling matches comparable models on certain benchmarks while using roughly a third of the tokens—which matters enormously when you're running a model millions of times inside automated workflows. The cost curve is the competitive surface, not the peak score.

There's also a "thinking effort dial" that lets developers sweep reasoning intensity between a low and high setting, trading speed for accuracy depending on the task. That kind of developer-facing control is relatively uncommon and reflects Thinking Machines' core thesis: that the people deploying AI into production need configurability more than they need a model that occasionally achieves a perfect score on an exam no human can pass.

On the safety side, Thinking Machines says it trained Inkling against an internal behavior specification across all modalities, with external red-teamers brought in to verify the results. For an open-weight model—one that anyone can download and run without oversight—that kind of documented safety work matters more than it does for a gated API.

One genuinely distinctive design choice: what Thinking Machines calls "epistemics." Rather than training Inkling to always produce a confident-sounding answer, they used reinforcement learning against proper scoring rules on resolved real-world questions, teaching the model to flag uncertainty and hedge appropriately. The video notes it uses a claims grader that runs web search to verify factual claims during training and penalizes unverified ones. Whether this holds under adversarial pressure in production is a different question—but the design intent is sound, and it's a meaningful departure from the confident-hallucination pattern that makes most LLMs exhausting to fact-check.

The Architecture Story Is the Geopolitical Story

Here's where things get uncomfortable in ways that extend well beyond model cards.

According to the Financial Times (cited in the video's source list), Inkling's MoE architecture largely follows DeepSeek V3—a Chinese model—and its post-training used synthetic data generated by Kimi K 2.5, a model from Moonshot AI, another Chinese lab. The video is careful to note that DeepSeek and Kimi's models are both open-weight and permissively licensed, so Thinking Machines did nothing wrong. Learning from published work is how science has always moved forward, in every field.

But the rhetorical double standard is hard to ignore. When DeepSeek released its models in early 2025, OpenAI and US officials accused it of distilling American models—the word "theft" was thrown around in congressional hearings. Now an American lab founded by one of OpenAI's most senior former executives has done something structurally similar in the other direction, and the framing has quietly shifted to "engineering."

The video puts it directly: "Now the arrow points the other way, and suddenly it's just engineering." That's not an accusation—it's an observation about how selectively the theft narrative gets applied.

There's a subtler signal buried here too. Murati knows OpenAI's technical stack as well as anyone alive. Given a blank slate, $2 billion, and access to Nvidia's top hardware, she built from DeepSeek blueprints and Kimi training data rather than anything resembling her former employer's approach. That's a data point about where she thinks the most productive foundations currently sit—and it arrives at a moment when the gap between Chinese open-weight models and American closed models has been visibly narrowing on the benchmarks developers actually use for hiring decisions.

The Business Model Is the Real Bet

Inkling is not the product. Tinker is the product. Inkling is the ingredient.

Thinking Machines isn't charging per token through a metered API. The revenue model runs through Tinker, the fine-tuning platform, with Inkling positioned as a permissively licensed base that enterprises customize and run on their own infrastructure. The logic is that a specialist model fine-tuned on your proprietary data can outperform a frontier generalist in your specific domain—at a fraction of the cost.

Futurum analyst Mitch Ashley, quoted in the video, called Inkling "the first credible Western alternative" to Chinese open models and flagged that picking a base model is effectively an architecture decision with compounding switching costs. Constellation Research analyst Holger Mueller argued the business model itself might be the real innovation: "charging for the platform instead of model access accelerates the commoditization of LLMs and makes AI ROI easier for businesses."

The timing is deliberate. Per Reuters (cited in the video's sources), American enterprises have been heavy users of Chinese open models, which are excellent and nearly free. Washington is now applying political pressure on that. The State Department has reportedly warned companies about Chinese model use, and congressional investigations into enterprise adoption have begun. For anyone with government contracts or regulatory exposure, Chinese models are becoming a liability regardless of their technical merit.

Meta's Llama family has also been a less dominant presence in the open-weight space following a disappointing Llama 4 release, which left a gap in the Western open ecosystem that Inkling is explicitly positioning to fill. Per TechCrunch's coverage, Thinking Machines is leaning into the "platform not a model" framing hard.

The uncomfortable question underneath all of this: if American policy successfully walls off the strongest open-weight foundations—which currently include Chinese models—while Western alternatives like Inkling lag on raw capability, what happens to derivative products built on each? A fine-tuned model inherits the ceiling of its base. The video raises this directly: companies in jurisdictions without restrictions, building on the strongest foundations available, may end up with better specialized tools than those who are policy-constrained to weaker bases. That would be a strange outcome for a technology policy designed to protect American competitive advantage.

Inkling is available on Tinker today. Whether it becomes the Western open-weight standard or a historical footnote in the AI consolidation story probably depends less on the model itself and more on whether Thinking Machines' bet—that the platform beats the model, and that "permitted" beats "powerful" in enough enterprise contexts—turns out to be right.

Dev Kapoor covers open source software and developer communities for Buzzrag.