Mira Murati left OpenAI in 2024. She spent the next year building Thinking Machines Lab. On July 15, that lab shipped its first model.

It's called Inkling. 975 billion total parameters, 41 billion active per token — a mixture-of-experts design, meaning the model routes each request through a fraction of its total capacity rather than engaging everything at once. Apache 2.0 license. Multimodal natively: text, images, and audio all in the same request. Context window of 1 million tokens.

The specs are solid. But there's a sentence in the announcement I keep coming back to: Thinking Machines says, explicitly, that Inkling is "not the strongest overall model available today, open or closed."

That's an unusual thing to put in a launch post.

Source spread

• Thinking Machines Lab — Introducing Inkling [hype] — The company's own framing: calibrated uncertainty, customization platform, explicit "not the strongest" positioning.
• TechCrunch — Thinking Machines' bet against one-size-fits-all AI [builder] — Covers the Tinker platform and why fine-tuning is the actual product.
• VentureBeat — Inkling, focused on low cost and 'resistance to censorship' [skeptic] — The censorship-resistance training angle and what Cognition's benchmark found.
• MarkTechPost — Inkling architecture deep dive [builder] — Technical breakdown, benchmark methodology, and the controllable thinking-effort feature.

What's real

• SWE-bench Verified at 77.6% is genuinely competitive. That's where the best coding models were two quarters ago. Not frontier today, but no embarrassment for a first release.
• AIME 2026 at 97.1% is striking. Most models cluster in the 70s on this math reasoning test. 97.1% means something real is happening in the reasoning stack.
• Audio natively, not bolted on. Text, images, and audio in the same request window. Most open-weight models handle text and image but require a separate pipeline for audio. Inkling doesn't.
• Apache 2.0 is the best possible license. No commercial restrictions, no scale thresholds, no fine-print. Download, modify, ship.
• Calibrated uncertainty is a real, trained-for property. The model flags when it doesn't know rather than guessing. That sounds like a small thing until you've watched a confident wrong answer from a different model get copy-pasted into a customer-facing document.

What deserves a side-eye

• The "not the strongest" claim is accurate, not false modesty. On Artificial Analysis' aggregate index, Inkling sits behind Fable 5, GPT-5.6 Sol, and Kimi K3. If raw benchmark score is your decision criterion, there are stronger options.
• Pricing changed July 17. The $1.87/$4.68 per million input/output was a 50% launch discount. OpenRouter offers $1/$4.05 and is the smarter starting point until you need Tinker's fine-tuning features.
• Training data undisclosed. We know it was 45 trillion tokens across text, image, audio, and video. We don't know the composition or copyright provenance. For production compliance use cases, that blank matters.
• The "resistance to censorship" training needs precision. Thinking Machines trained Inkling to engage with politically or institutionally sensitive questions rather than auto-refusing them. They also trained it to refuse weapons, CBRN, manipulation, and abuse requests — those two things aren't in conflict. But "censorship non-compliance" as a named feature will be used creatively by some downstream developers, and Thinking Machines will have to police that at the application layer since they weakened the guard at the model layer.

97.1%

→ Source: Thinking Machines Lab

What to do about it

• If you're evaluating for coding agents: Inkling's 77.6% SWE-bench Verified is real and worth testing. Apache 2.0 means no legal friction. Start with OpenRouter at $1/$4.05 before committing to the Tinker platform pricing.
• If you need native audio input: This is one of the few open-weight models that handles text, image, and audio in a single request — no separate transcription pipeline. Test whether the consolidated architecture reduces your latency before assuming it's always cleaner.
• If you're deploying in a regulated context: Training data provenance is undisclosed. Run your compliance team's checklist before building a production system on top of it.
• For everyday users: Inkling will surface inside tools built on it, not usually directly. If you're curious what an AI that says "I don't know" instead of guessing actually feels like, the model card at thinkingmachines/Inkling is worth reading — it's the clearest public explanation of why calibrated uncertainty is a feature, not a limitation.

Further reading

• Thinking Machines — Introducing Inkling — primary source
• TechCrunch — Thinking Machines' first model — full coverage
• VentureBeat — Inkling, resistance to censorship — the censorship-training angle
• HuggingFace — thinkingmachines/Inkling — weights and model card
• OpenRouter — Inkling pricing — API access at launch prices