Murati Ships at Last: Inkling Is the Largest US Open-Weight Model, and It Isn't Trying to Win

Seventeen months, one model, zero swagger

Thinking Machines Lab spent a year and a half as the most-watched empty room in AI. Founded in February 2025 by former OpenAI CTO Mira Murati, it raised $2 billion at a $12 billion valuation before shipping anything at all, hired aggressively, and then mostly said nothing. On July 15 it finally released a model.

The model is called Inkling. It is a mixture-of-experts transformer with 975 billion total parameters and roughly 41 billion active per token, a context window of up to one million tokens, and pretraining across 45 trillion tokens of text, images, audio and video. The full weights are on Hugging Face under Apache 2.0 — which, by parameter count, makes it the largest US-built open-weight model publicly available. A lighter sibling, Inkling-Small, ships as a preview at 276B total and 12B active parameters, with full weights promised once testing wraps.

The unusual part isn't the spec sheet. It's the framing. Thinking Machines' own announcement states plainly that Inkling is "not the strongest overall model available today, open or closed." A lab that has been sitting on more expectation than almost any startup in the field opened its first release by telling you it did not win.

What the numbers actually say

The company published benchmarks at maximum reasoning effort (what it labels effort=0.99): 29.7% on Humanity's Last Exam text-only, 97.1% on AIME 2026, 87.2% on GPQA Diamond, 77.6% on SWEBench Verified, 63.8% on Terminal Bench 2.1, 73.5% on MMMU Pro vision, and 91.4% on VoiceBench. Adversarial safety on the FORTRESS eval came in at 78.0%.

Read those as competitive-tier rather than frontier-defining. They put Inkling in serious company without displacing anyone at the top — consistent with what the lab says about itself. The more interesting claims are about efficiency: Thinking Machines says Inkling reaches equivalent coding performance while spending roughly a third as many tokens as Nvidia's Nemotron 3 Ultra. In a market where inference cost is the binding constraint on agentic deployment, tokens-to-solution is arguably a more honest metric than a leaderboard row.

The architecture notes are worth flagging for anyone building on top of this. Training used a hybrid optimizer — Muon for large matrix weights, Adam for the rest — and an attention stack mixing sliding-window and global layers at a 5:1 ratio. Post-training scaled past 30 million rollouts, with the lab reporting log-linear reasoning gains from reinforcement learning. That last claim is the kind of thing that either replicates in the open-weights community within weeks or quietly doesn't, and now that the weights are public, we'll find out.

The business is Tinker, not the model

Inkling is not the product. Thinking Machines does not monetize it. Revenue comes from Tinker, the company's fine-tuning platform, where Inkling went live for customization on day one with a 50% introductory discount. The weights are the customer-acquisition funnel; the customization layer is the toll booth.

That inverts the standard frontier-lab arrangement, where the model is a locked API and the margin sits on per-token inference. Thinking Machines is instead betting that enterprises want a strong, malleable base more than they want the single best general model. The showcase evidence is a Bridgewater Associates partnership: a model trained on the hedge fund's financial expertise scored 84.7% on financial reasoning tests at roughly a fourteenth the cost to run. That number deserves an asterisk — as TechCrunch notes, it comes from the companies' own evaluation, not an independent one.

The distribution push was immediate and broad: Together AI, Fireworks, Modal, Baseten, and Databricks all lit up support. Databricks is pitching it through its Unity AI Gateway with the exact argument the open-weights camp has been making for two years — fine-tune on proprietary codebases, plug into coding agents like Cursor and OpenCode, "optimize inference spend without per-token API pricing," avoid vendor lock-in.

The censorship angle is a positioning move

Thinking Machines evaluated Inkling on Cognition's Propaganda and Censorship Eval and reported strong patterns of censorship non-compliance. That is not an idle capability note. It is a direct pitch at the anxiety that has been building all year as Chinese open-weight models — DeepSeek, Qwen, GLM, and now Moonshot's Kimi K3 — captured a growing share of open-model deployment on cost grounds.

Fortune frames the strategic gap clearly: Meta pivoted away from open weights toward paid proprietary offerings, OpenAI's open releases stayed limited, and cost-sensitive businesses drifted toward Chinese alternatives, which is increasingly read in Washington as a national concern. Inkling is the first credible US answer at genuine scale. Whether "resistance to censorship" survives contact with a fine-tuned derivative is a different question — the whole point of open weights is that anyone can retrain the behavior out.

Where the hype needs trimming

Three caveats matter.

First, the company acknowledges using competitor outputs for post-training via distillation, saying future models will rely on self-contained methods. That's an honest disclosure, and also a reminder that "independent frontier lab" and "trained partly on other labs' outputs" are not mutually exclusive.

Second, the finances are murky. The reported $50 billion funding round had stalled by January, and two co-founders left for OpenAI in the same month. A lab burning capital at ~200 employees while giving away its flagship needs Tinker to convert, and fast.

Third, Apache 2.0 on 975B parameters is more symbolic than practical for most organizations. Very few can actually serve a model this size themselves — which is precisely why the deployment-partner list matters more than the license. In practice, most users will rent Inkling from Databricks or Fireworks rather than run it. Open weights buy portability and audit rights, not free inference.

The takeaway

Inkling is the most interesting model release of the week not because it tops anything, but because of what it refuses to claim. A lab with $12 billion of expectation behind it shipped a model, said out loud that it isn't the best, and pointed customers at the fine-tuning layer instead. That's either admirable strategic clarity or an elegant way to lower the bar before jumping — and the honest answer is that we won't know until we see whether enterprises actually pay for customization at volume.

What's not ambiguous: the United States now has a large, permissively licensed, genuinely multimodal open-weight model with real deployment infrastructure behind it on day one. For eighteen months the open-weights conversation has been drifting toward Hangzhou and Beijing. Inkling doesn't reverse that, but it's the first American release in a while that makes the argument on the merits rather than the flag.