Murati Shipped a Frontier You Can Actually Own

Mira Murati just shipped the largest American open-weights frontier model — and she bet enterprises care more about owning the checkpoint than topping the leaderboard.

What's new: On July 15, Thinking Machines Lab released Inkling — a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, native text/image/audio inputs, up to 1 million tokens of context, and full weights under Apache 2.0 on Hugging Face. The company trained it from scratch on Nvidia GB300 systems over 45 trillion multimodal tokens in roughly nine months with ~200 employees. APIs are live on Tinker, TogetherAI, Fireworks, Modal, Databricks, and Baseten.

Why it matters: June's Anthropic export ban proved Washington can flip proprietary APIs off in 90 minutes. Inkling is the structural opposite — a checkpoint you download, fine-tune, and host. That is the product thesis Murati has been funding since a record $2 billion seed round at a $12 billion valuation in July 2025, before she had shipped anything.

The bet:

• Thinking Machines admits Inkling is "not the strongest overall model available today, open or closed" — and ships anyway
• The pitch is customization: controllable "thinking effort" from 0.2 to 0.99, fine-tuning on Tinker, multimodal agent workflows
• 77.6% on SWE-bench Verified beats U.S. open rival Nemotron 3 Ultra's 70.7%, per the company's published benchmarks
• 1257 on Design Arena's Agentic Web Dev leaderboard — competitive with closed models, strongest among open weights in that human eval
• Chinese rivals still lead on several hard benchmarks: GLM 5.2 posts 82.7 on Terminal Bench 2.1 vs Inkling's 63.8

The vacuum Inkling fills:

• Meta's last open-weight release was Llama 4 in April 2025; Muse Spark launched April 8, 2026 as a proprietary API with no downloadable weights, per Artificial Intelligence News
• Alexandr Wang told developers bigger models are coming and "plans to open-source future versions" — no date, no weights tonight
• Western builders who refuse Chinese defaults have been choosing between Nemotron-scale U.S. open models and closed subscriptions
• The Register, July 16: Inkling is "the largest American open weights model to date"

The policy backdrop:

• Convina's July 15 reporting flagged White House talk of new action on open-source models — the same week Inkling landed on Hugging Face
• Closed labs sit behind API contracts Commerce can pressure; downloaded weights do not phone home for permission
• Axios notes Inkling's final training phase used data from existing open models including Moonshot's Kimi K2.5 — a reminder that "sovereign" stacks still inherit global open-weight supply chains

The quotes:

• Thinking Machines blog, July 15: "Our mission is to build AI that extends human will and judgment" — Inkling is the first model released with full weights "so that people can make it their own"
• Co-founder John Schulman, on X, July 15: "Inkling is out today, with open weights and in Tinker. It's been fun to watch this one come together: pretraining began last winter, and starting in mid-January a small team built up the coding, reasoning, and agentic training from there."
• Axios, July 15: Murati is betting enterprises "care less about the smartest general-purpose model than one they can make their own"
• The Register, July 16: Murati's lab aims to change a market where "your options are few and far between outside of the Chinese model houses"

What we cannot verify: Independent third-party audits of Thinking Machines' benchmark tables are not yet published. How many enterprises will actually run a model requiring eight Nvidia B300s at native precision is untested — an NVFP4 quantized checkpoint cuts hardware roughly in half. Revenue impact on closed labs won't show up for quarters. Whether Washington's open-source scrutiny targets weights Murati just published is still rumor, not rule text.

Convina's view: Murati did not ship the smartest model on the planet. She shipped the smartest procurement argument of the summer: a frontier-class checkpoint Western enterprises can fork, fine-tune, and air-gap while Meta sells APIs and Washington practices kill switches on closed endpoints. Benchmark crown chasers will keep renting Claude and GPT. Everyone who lived through June's Mythos blackout just got a $12 billion-funded alternative that does not ask permission to stay online. The open-weights war was China's to lose until an American lab stopped apologizing for second place and mailed the file anyway.