Thinking Machines Ships Inkling, Its First Open-Weight Model — and Says It Isn't the Best

Most AI labs launch a model by telling you it's the best one. Thinking Machines Lab launched its first model on Wednesday by telling you it isn't.

"Inkling is not the strongest overall model available today, open or closed," the company wrote in its announcement — a sentence sitting a third of the way into a post otherwise dense with benchmark tables and reinforcement-learning curves. It is not a hedge buried in a footnote. It is the pitch.

Inkling, released July 15, is the first model from the startup Mira Murati founded in February 2025 after leaving OpenAI, where she was CTO. It is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token, pretrained on 45 trillion tokens of text, images, audio, and video. It supports up to a 1 million-token context window, reasons natively over text, images, and audio, and ships with full weights on Hugging Face under an Apache 2.0 license — the permissive kind, no field-of-use restrictions, no revenue thresholds. Alongside it, Thinking Machines previewed Inkling-Small, a 276B-parameter sibling with 12B active, whose weights it says it will release once safety testing wraps.

What's actually in the box

The architecture is a 66-layer decoder-only transformer whose MoE design, per the company, "largely follows DeepSeek-V3": 256 routed experts plus 2 shared experts per layer, with 6 routed experts firing per token. The departures are real, if wonky. Thinking Machines interleaves sliding-window and global attention layers at a 5:1 ratio and encodes position with relative positional embeddings rather than the near-universal RoPE, claiming better extrapolation to long sequences. Audio goes in as dMel spectrograms and images as 40x40 pixel patches — no separate encoder, consistent with the "interaction models" system previewed in May. Training used a hybrid optimizer (Muon for large matrix weights, Adam for the rest) on NVIDIA GB300 NVL72 systems, and post-training scaled past 30 million RL rollouts.

The benchmarks bear out the modesty. On Humanity's Last Exam (text only), Inkling scores 29.7% against Claude Fable 5's 53.3% and GPT 5.6 Sol's 47.2%. On SWEBench Verified it hits 77.6%, versus 95.0% for Fable 5. It lands mid-pack even among open-weight peers, trailing GLM 5.2 on both HLE and Terminal Bench 2.1. It does lead its class on a few axes — MCP Atlas (74.1%) and FORTRESS adversarial safety refusals (78.0%, the best open-weight score in the company's comparison set) — and posts 97.1% on AIME 2026 and 87.2% on GPQA Diamond.

The number Thinking Machines actually wants you to look at is a ratio. Inkling has a controllable "thinking effort" dial, and swept across it, the model matches Nvidia's Nemotron 3 Ultra on Terminal Bench 2.1 while spending roughly a third as many tokens. For a model you run millions of times inside an agent loop, that curve — not the peak score — is the product.

There's also a candid disclosure most labs would bury: to bootstrap post-training, Thinking Machines ran an initial SFT pass on synthetic data generated by other open-weight models, including Moonshot AI's Kimi K2.5. It says this was a small fraction of compute, and that the next model will use fully self-contained post-training.

The bet

Here is the strategic knot. Apache 2.0 means nobody who downloads Inkling owes the company anything — no metered tokens, no API contract, no seat licenses. Fortune, citing Bloomberg, reports the startup isn't trying to monetize the model at all. Revenue has to come from Tinker, its fine-tuning platform, where Inkling is available today at 64K and 256K context options at a limited-time 50% discount, plus a slice of the hosting ecosystem it stitched together across Together AI, Fireworks, Modal, Databricks, and Baseten.

The thesis: a frontier model rented from a central lab is the same product everyone else rents, refined by someone else, on someone else's schedule. A slightly weaker model you own and shape with your own institutional knowledge beats it on your actual work. The company's clearest evidence is a joint project with Bridgewater Associates — not an investor — where researchers fine-tuned an existing open-source model on the hedge fund's financial expertise and scored 84.7% on financial reasoning tests, beating top proprietary models at roughly a fourteenth of the run cost. That result comes from the two companies' own evaluation, not an independent one.

The argument has unusually well-placed backup this week. In a July 12 blog post, Microsoft CEO Satya Nadella — whose company has poured billions into both OpenAI and Anthropic — warned enterprises they are paying for intelligence twice. Hugging Face CEO Clem Delangue told TechCrunch last week that frontier models will increasingly be reserved for experimentation and high-value tasks while production work shifts to open alternatives. Open models already accounted for 29% of traffic through Vercel's gateway last month.

Why it might not work: fine-tuning requires serious ML talent that most enterprises don't have, and Tinker's whole premise is that the customer takes on responsibility for making customizations safe — a burden OpenAI and Anthropic absorb on their customers' behalf. Apache 2.0 also means the ecosystem Thinking Machines is cultivating can route around it entirely; SGLang, vLLM, and llama.cpp support landed on day one, which is great for adoption and terrible for capture. And "not the strongest model" is a claim that ages in only one direction if the next release doesn't close ground.

What to watch

Three things. Whether Inkling-Small's full weights ship on schedule — a 12B-active model that nearly matches its 41B-active sibling on reasoning is arguably the more commercially interesting artifact. Whether Tinker converts downloads into paying fine-tuning runs, the only mechanism by which any of this becomes a business. And the money: Thinking Machines raised $2 billion at a $12 billion valuation, a reported $50 billion round was said to have stalled in January, and Nvidia's March partnership for a gigawatt of Vera Rubin capacity came with what Nvidia called a "significant investment." The company employs roughly 200 people. Its bet may be less that it will eventually spend like OpenAI than that it won't have to.