"It isn't the strongest model" — who introduces their debut like that?

On July 15, the AI world got hit with two strange sentences at the same time. One: "Mira Murati's company finally shipped its first model." The other came straight from the company itself: "Inkling is not the strongest overall model available today, open or closed." Startups usually dress up their debut like the greatest thing ever built. This one did the opposite. It nailed "we're not the best" right onto the front door.

The reason this shakes things up is that it isn't a typo or false modesty — it's the strategy. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, didn't chase a benchmark crown with its first model. Instead it placed its whole bet on a single idea: AI that a company can rip open and reshape itself will, in the end, beat the one-size-fits-all models the big labs sell. It's an attempt to move the axis of the game from a raw-performance contest to a customization contest.

And spec-wise, this is no small thing. Inkling is a 975B-parameter mixture-of-experts (MoE) model, but it only activates roughly 41B of those parameters for any given task. It handles text, images, audio, and video natively as four modalities, was trained on 45 trillion tokens, and takes a context window of up to 1 million tokens. Those are frontier-class numbers — and yet the company declared it isn't selling "the strongest." Let's unpack why they made such a weird-looking bet.

Who's behind this — Murati, Thinking Machines, and the open-weight camp

Mira Murati carries serious weight in this industry. As OpenAI's CTO, she oversaw the development of huge products like ChatGPT, GPT-4, DALL-E, and Sora. She left OpenAI in 2025 and founded Thinking Machines Lab, staffing it heavily with fellow OpenAI alumni. The industry read it as "a challenger carrying OpenAI's DNA."

It isn't just the people that make the company formidable — the money made headlines too. Per TechCrunch, a fundraising round reportedly in the neighborhood of $50 billion was in play (it moved between last November and January, stalled for a bit, and the company hasn't disclosed its current funding status). In March it partnered with Nvidia to secure compute capacity. In other words, this was a company with the money, the people, and the GPUs — but no product yet. Now it has finally put its first in-house model into the world.

Here's the concept you need: "open-weight." Frontier models like ChatGPT, Claude, and Gemini are API-only. They live locked inside the vendor's servers; you send a request and get an answer back, but you never touch the model's internals (its weights). An open-weight model is the opposite — you can download the weight files themselves, load them on your own servers, and rework (fine-tune) them with your own data. Inkling is the latter. The full weights actually shipped on Hugging Face, released under a commercially permissive Apache 2.0-style license.

The core frame Murati threw down is this: expertise is usually domain-specific, and that know-how is irreplaceable, so AI an organization shapes for itself beats a centrally stamped, one-size-fits-all model. Put simply, the future isn't selling every company the same do-everything model — it's selling each company a good raw stone they can carve to fit their own data and workflows. Inkling is that stone.

The core details — a design obsessed with efficiency

Start with the architecture. Inkling is a 66-layer decoder-only transformer with a sparse MoE backbone. Each MoE layer holds 256 routed experts plus 2 shared experts. When a token comes in, only 6 routed experts and 2 shared experts fire. That's how you get 975B total parameters while only 41B actually run at once. "Enormous body, mid-size compute bill" is exactly the appeal of MoE.

The details are all in on efficiency too. Attention alternates sliding-window and global layers at a 5:1 ratio and uses just 8 KV heads to save memory. For positional information it swapped today's standard RoPE for relative positional embeddings, and it slots short convolutions into the attention layers — several experimental choices that break from the crowd. Training optimization was a hybrid: Muon for the large matrix weights, Adam for the rest. And post-training ran over 30 million reinforcement-learning rollouts.

So how efficient is it? The company claims it uses about a third as many tokens as Nvidia's Nemotron 3 Ultra to reach equivalent coding performance. Not the benchmark champion, but clearly aiming to win on bang-for-buck. One honest caveat: even though it trained on 45 trillion tokens spanning text, image, audio, and video, current output is text-only. It reads and reasons across four modalities, but it only generates text for now.

Worth nailing down the distribution too. Inkling is served on the company's own fine-tuning platform, Tinker, with 64K and 256K context options, plus a limited-time 50% discount. It's also wired into deployment partners like Together AI, Fireworks, Modal, Databricks, and Baseten so you can reach it through multiple paths. And it previewed a smaller sibling: Inkling-Small, a 276B-parameter MoE with 12B active, which the company says matches or beats its big brother on many benchmarks. Those weights land once testing wraps.

Item Inkling Inkling-Small Closed frontier (GPT/Claude/Gemini)
Total parameters 975B MoE 276B MoE Undisclosed
Active per task ~41B ~12B Undisclosed
Modalities (input) text, image, audio, video multimodal mostly multimodal
Context up to 1M tokens large varies by model
Training tokens 45T not disclosed Undisclosed
Distribution open weights (download, fine-tune) open weights (planned) API-only (weights closed)

Who gains what — what a company actually gets from this

The biggest winner is any company that wants an AI of its own. Until now most firms rented OpenAI's or Anthropic's API. Convenient, sure, but running someone else's model on someone else's servers always dragged three problems along. One: the security and compliance burden of sending your sensitive data to an outside server. Two: subscription lock-in with per-token billing bleeding out forever. Three: the model never truly, deeply knows your domain (the specialized jargon of law, medicine, finance). An open-weight model solves all three at once — you put the weights on your own servers, fine-tune with your own data, and run it inside your own infrastructure with no subscription fee.

There's evidence this isn't hot air. Per TechCrunch, in a collaboration with hedge fund Bridgewater, an open-source model fine-tuned on financial expertise scored 84.7% on financial-reasoning tests — beating several commercial closed models while costing roughly one-fourteenth as much to run (worth noting these are the companies' own evaluation metrics). The point is clear: even if general performance runs a bit lower, carving a model to your domain can make it cheaper and more accurate than the strongest model out there. That's exactly the spot Murati's bet is aiming at.

Thinking Machines itself makes money off this structure. Its revenue model isn't "model usage fees" (per-token billing) — it's Tinker, the customization platform. Release the weights for free, then earn from the tools and infrastructure that make fine-tuning easy. It's a classic "open-core" play: grow an ecosystem with open source, then monetize services on top. So giving the weights away freely isn't charity — it's the business plan.

Precedents — open vs. closed, the fork between success and failure

This "open vs. closed" fight isn't new. The open camp's flagship success is Meta's Llama. By effectively opening its large-language-model weights, Meta let developers and companies worldwide build countless derivatives on top — and Llama became the de facto standard base of the open ecosystem, cementing Meta's position as the ground where AI grows. France's Mistral is similar: relatively small but efficient open models that carved out a clear presence in Europe and the enterprise market. "Open the weights and the community grows it for you" is this camp's core logic.

On the other side, the closed camp's success is unquestionably OpenAI's GPT line. It kept the weights tightly wrapped, sold only through an API, and still dominated the market on overwhelming performance and speed to product. "Monopolize the best performance, open only the front door, and people will line up" — that strategy worked. Anthropic's Claude and Google's Gemini hold the frontier on the same closed path. So on the scoreboard so far, the closed camp owned "top performance" while the open camp owned "ecosystem reach."

There are failure lessons too. Going open doesn't automatically win. If performance is mediocre, if the fine-tuning tools and docs are thin, or if the license is murky, nobody uses your open weights no matter how freely you release them. And even a closed model, once it starts losing on performance, triggers instant "why am I paying for this expensive API?" churn. So Inkling's battleground is precise: not being the top performer, but answering one question — is this the best raw stone to fine-tune? By leading with Tinker and customization instead of benchmarks, Murati is aiming for a third seat, somewhere between Llama's reach and GPT's polish.

Competitor counter-plays — how do OpenAI, Anthropic, and Mistral respond?

OpenAI and Anthropic (the closed-frontier camp) counter by crushing on the performance gap. Now that Inkling has admitted "we're not the strongest," the closed camp can play the "our best model is still overwhelmingly better" card. Plenty of companies would rather just use an API that works than eat the hassle of fine-tuning. But there's a crack in that defense: even OpenAI has recently started shipping its own open-weight models to check the open camp — a tacit admission that "closed alone can't fully block customization demand."

Mistral and Meta's Llama (rivals inside the open camp) counter with the first-mover lock they already hold on the open ecosystem. A company that wants to fine-tune needs a reason to pick a newcomer like Inkling — and Llama and Mistral already sit on vast tooling, tutorials, and community assets. To fight that, Inkling's fine-tuning experience through Tinker has to be dramatically easier and cheaper. In other words, Inkling's real rival may not be GPT at all, but the other open models already dug in.

Hyperscalers and clouds (the neutral zone) will happily burn either fuel. Inkling launched already loaded onto deployment partners like Together AI, Fireworks, Databricks, and Baseten. For a cloud, open or closed doesn't matter as long as it runs on their infra and rings the register, so the open camp's spread is actually welcome news. In this setup, Thinking Machines' moat isn't the model itself — it's Tinker, the customization layer. Even if others just grab the Inkling weights and run, as long as Inkling owns "the easiest fine-tuning experience," the money keeps flowing back to the company. That's the play.

So what actually changes

If you're a developer or engineer — the choice between "rent a frontier API" and "fine-tune open weights" just got a lot more real and a lot heavier. Inkling stacks large-model performance potential (975B/41B) on top of open weights, a 1M context, and multimodal input, and claims Tinker lowers the fine-tuning difficulty on top. It's worth seriously considering a hybrid design when you architect AI systems: fine-tuned open model for your core domain, closed API for general grunt work.

If you're a business decision-maker — the central question is shifting from "which model is smartest?" to "can I rework it to fit my data, my regulations, my costs?" If the Bridgewater case (domain fine-tuning beating commercial models at one-fourteenth the cost) holds up, open weights may actually be the right answer for finance, medicine, law, and public-sector work handling sensitive data. Just weigh coldly that you carry the fine-tuning and operations headcount and infra costs — and that most performance figures are self-reported.

If you're an investor — the real message of this launch isn't "model performance," it's the business model. Thinking Machines declared it will earn from the Tinker customization platform rather than per-token billing, a signal that AI's revenue structure is splitting from "selling models" toward "customization and operations services." That said, weigh the risks: a company once talked about at a ~$50B valuation still hasn't shown concrete revenue, and how it covers its Nvidia compute bill remains murky.

If you're a general user — you're unlikely to touch Inkling directly anytime soon. But if this spreads, the AI running behind your apps and services may stop being "one do-everything GPT" and become "a dedicated model carved to that company's work." Your banking app runs a finance-tuned AI, your hospital app a medicine-tuned one. We may be at the mouth of a shift from the age of uniform chatbots to an age of custom AI with a different personality for every service.

🥄 Three Things You're Probably Wondering

— Is admitting "it's not the strongest" confidence or self-sabotage? Closer to strategic confidence. GPT, Claude, and Gemini are already bleeding each other dry in the benchmark-crown fight, so jumping in there offers poor odds. Instead they're going for first place in a new ring — "raw stone for fine-tuning." The catch: Llama and Mistral already staked that ring, so whether the admission works comes down to how genuinely convenient Tinker turns out to be.

— If the weights are open and free, how does the company make money? Open the weights, but monetize Tinker — the platform that makes fine-tuning and operating those weights easy. It's an "open-core" model, similar to how Red Hat made money on services even though Linux is free. The question is whether the open-model fine-tuning market is big enough to carry that revenue, and that's still unproven.

— It's trained on 45T tokens and multimodal, so why is output text-only? Right now Inkling reads and reasons across text, image, audio, and video, but only generates text. Understanding four input modalities and actually producing images, audio, or video as output are wildly different difficulty levels, so they're opening it in stages. Whether output modalities expand later is the next thing to watch on this model.

References

Numbers are as of announcement and may change.