The Open-Weight Bet: Inside Mira Murati's First Model, and Why It Refuses to Race

On the morning of July 15, 2026, a 975-billion-parameter model went up on Hugging Face with an Apache 2.0 license and a shrug. Thinking Machines Lab, the startup Mira Murati founded after leaving OpenAI's top technical seat, had just shipped its first product to the public — the largest open-weight model ever built on American soil — and led with an admission almost no AI lab makes about its own work: this is not the strongest model available today, open or closed.

That sentence is doing a lot of work. It arrived in the middle of what industry trackers are already calling the most concentrated ten days of model releases in AI history: xAI's Grok 4.5 on July 8, Anthropic's Claude Sonnet 5 at the end of June, OpenAI's three-tier GPT-5.6 family rolling out through the month, Google scrambling to rebuild Gemini from scratch, and Moonshot's Kimi K3 landing the day before Inkling. Nearly every major lab put something new on the table within the same stretch of weeks. In a market where "best model wins" had been the only story that mattered for three years, Thinking Machines chose to compete on a different axis entirely — and the reasoning behind that choice says more about where frontier AI is actually heading than another leaderboard score would.

Who is building this, and why it matters

Murati spent six years at OpenAI, joining in 2018 and becoming chief technology officer in 2022, where she oversaw the research, product, and safety work behind ChatGPT, DALL-E, Codex, and Sora. She stepped down in September 2024 to, in her words, do her own exploration. Four months later she launched Thinking Machines Lab as a public benefit corporation, with a stated mission to make AI systems more widely understood, customizable, and generally capable — a notably different framing than "build the smartest possible model," which has been the north star at OpenAI, Anthropic, and Google DeepMind alike.

She did not build it alone or quietly. Thinking Machines pulled in roughly thirty researchers and engineers from Meta, Mistral, and OpenAI itself, including OpenAI co-founder John Schulman and advisors Alec Radford and Bob McGrew — a roster senior enough that the seed round priced itself accordingly. Andreessen Horowitz led a $2 billion raise in July 2025 at a $12 billion valuation, with Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street all participating. By November, Bloomberg reported the company was in talks for a follow-on round at $50 to 60 billion — nearly five times the seed price in under five months, though that round has not yet closed.

What Inkling actually is

Inkling is a mixture-of-experts model: 975 billion parameters in total, but only about 41 billion of them switch on for any given request, which is what keeps inference costs manageable at that scale. It was trained on roughly 45 trillion tokens spanning text, image, audio, and video, and it reasons natively across all four rather than bolting on separate encoders after the fact. Full weights are posted on Hugging Face under an Apache 2.0 license, meaning any developer, university lab, or competitor can download the entire model, inspect it, fine-tune it, and redeploy it without asking permission or paying a licensing fee. Same-day fine-tuning support ships through Tinker, the company's developer product for customizing models — already in use at customers including the hedge fund Bridgewater Associates.

That business model is the quiet part of the story. Thinking Machines isn't primarily trying to sell access to Inkling the way OpenAI sells ChatGPT subscriptions or Anthropic sells API credits. It's trying to sell the tooling that lets other people bend an open model into something specific to them — a bet that the money in the next phase of AI is in customization infrastructure, not in owning the single smartest model on the planet.

The open-weight gap it's trying to close

Here is the part of the story that has less to do with Silicon Valley bravado and more to do with geography. For the past two years, the most capable open-weight models in the world have not come from the United States. They've come from China — DeepSeek, Alibaba's Qwen line, and Moonshot's Kimi models have dominated open leaderboards, downloaded by researchers and startups worldwide who can't afford or don't want closed API access. American labs, by contrast, kept their frontier work locked behind APIs: OpenAI, Anthropic, and Google all built closed systems, ceding the open-source conversation almost entirely to Chinese developers.

Inkling is, bluntly, the largest open-weight model an American lab has ever released — and it arrives at a moment when Washington is actively worried about Chinese labs extracting capability from closed U.S. models through techniques like distillation. An open American alternative changes that conversation, whether or not that was the explicit intent.

Reporting on Inkling's technical lineage has noted it incorporates techniques and components with roots in DeepSeek and Kimi's own published research — a small irony worth sitting with. The frontier is no longer a straight line from any single country outward; it's a shared, cross-pollinated ecosystem where American, Chinese, and now increasingly European and Middle Eastern labs are all building on each other's published work, openly, in ways that make "whose model is ahead" a much blurrier question than it was even a year ago.

Why "not the strongest" might be the smarter play

The obvious objection to Inkling's strategy is that customers want the best model, full stop. But that logic assumes a single global market for "the best," and 2026's frontier landscape increasingly rejects that assumption. OpenAI split GPT-5.6 into three tiers — Sol for heavy reasoning, Terra for mid-cost quality, Luna for cheap high-volume work — priced from $1 to $30 per million tokens, explicitly because no single model configuration serves a hedge fund's risk desk and a customer-support chatbot equally well. Google scrapped a nearly finished Gemini successor and restarted from scratch rather than ship something merely incremental. The era of "one model to rule every use case" is ending industry-wide, and Inkling's bet — give developers a very good, fully modifiable foundation and sell them the tools to shape it — is a coherent answer to that same shift, just from the open-source side of the ledger rather than the tiered-pricing side.

The infographic: ten days, five labs

Where the money actually is

It is worth dwelling on the Tinker piece a moment longer, because it is the clearest signal of how Thinking Machines expects to make money in a market that already has three well-funded incumbents selling frontier intelligence directly. Rather than compete on the API-access model that OpenAI, Anthropic, and Google have all built their businesses around, Tinker sells the harder, more specialized layer underneath: the tooling to actually fine-tune a large model on a company's own data, safely, without needing an in-house machine learning infrastructure team to do it. Bridgewater Associates, one of the world's largest hedge funds, is reportedly already a customer — a sign that the pitch is landing with exactly the kind of technically sophisticated, data-rich enterprise buyer who has the most to gain from a model they can shape privately rather than query through someone else's API.

That positioning also explains why Inkling didn't need to top any leaderboard to be a meaningful release. A hedge fund fine-tuning a private trading-research assistant, a hospital system building a clinical-documentation tool on data that legally cannot leave its own servers, or a university lab probing a model's internals for interpretability research all have reasons to prefer an open, modifiable 975-billion-parameter foundation over a closed system that scores three points higher on a public benchmark. Thinking Machines is betting there are enough of those buyers to build a durable business without ever winning the capability race outright — a genuinely different theory of how value gets created in frontier AI than the one every other major lab is currently running.

What this means for the rest of us

For developers and smaller companies, an open, Apache-licensed model at this scale is a meaningful cost lever: no per-token API bill, no vendor lock-in, the ability to run it on your own infrastructure or fine-tune it privately without your data ever leaving your servers. For researchers, an open weight file at 975 billion parameters is a genuine research object — something that can be probed, interpreted, and studied in ways closed APIs never permit. For safety researchers, that openness cuts the other way: once weights are public, there is no way to revoke access, patch a discovered vulnerability at the source, or stop a bad actor from fine-tuning away whatever guardrails shipped with the base model. Every open-weight release is a one-way door.

That tension — openness as public good versus openness as unrecoverable risk — is not new, but Inkling raises the stakes simply by being the biggest American entry in that category to date. It will not be the last. If the pattern holds, expect more labs to split their strategy the way Thinking Machines has: a closed, cutting-edge system for enterprise customers who need the very latest capability, and an open, slightly-behind-the-frontier release for the researchers, students, and smaller builders who make up the actual texture of the AI ecosystem. The race for the single smartest model continues elsewhere. Inkling is a bet that it was never the only race worth running.

Sources