Inkling — the first open-weights model from Thinking Machines Lab, the startup founded by Mira Murati, former CTO of OpenAI

AI Dev Tools — General
, ,
1

Inkling
Inkling1536×1024 286 KB

By Grego · Category: AI & LLMs

There are releases that matter because of the model, and there are releases that matter because of the thesis underneath. Inkling — the first open-weights model from Thinking Machines Lab, the startup founded by Mira Murati, OpenAI’s former CTO — is one of the latter. Yes, it’s a multimodal MoE with 975B total parameters and 41B active, 1M token context, and full weights downloadable from Hugging Face. But what we should be discussing in Latin America isn’t the size: it’s that Thinking Machines released it under the Apache-2.0 license and said, explicitly, that “it is not the strongest model available today”. That combination — opening the weights and explicitly renouncing the fight for maximum score — is a strategic bet, not a marketing oversight.

And for teams in the region, that bet changes the calculus of which model to choose.

What is Inkling, in a nutshell

Before the thesis, the facts. Inkling is a 66-layer decoder-only transformer with Mixture-of-Experts backbone (256 routed experts + 2 shared), 975B total parameters and 41B active per token. It was pretrained on 45 trillion tokens of text, images, audio and video, with an encoder-free multimodal architecture. It supports up to 1M context tokens and was released with an NVFP4 checkpoint for efficient inference. The weights are on Hugging Face (thinkingmachines/inkling) and the model is available for fine-tuning from day one on Tinker, Thinking Machines’ own customization platform, plus several inference partners (Together, Fireworks, Modal, Databricks, Baseten).

There’s also a preview of Inkling-Small (276B total, 12B active), designed for cost and latency-sensitive workloads, still in testing.

None of this is the interesting part. The interesting part is why they did it this way.

The thesis: your advantage isn’t in the model, it’s in what you add to it

Thinking Machines’ argument is uncomfortable for the dominant business model: they argue that organizational knowledge is specific to whoever has it, and a central lab selling you a uniform product will almost always perform worse than a version you’ve fine-tuned on your own data. In their framework, Inkling isn’t a consumer product competing with a chatbot; it’s a customization base for enterprises, and Thinking Machines’ business isn’t subscriptions but Tinker: training, fine-tuning, and hosting.

It’s not a lone voice. Satya Nadella has been warning that companies “pay twice” with closed models — the subscription, plus the business knowledge you hand over to the vendor every time you use it. The CEO of Hugging Face has predicted that serious production work will migrate to private or open-source alternatives. Inkling is that thesis converted into downloadable weights under Apache-2.0.

For a dev in Buenos Aires, Bogotá, or Santiago, what matters is the license. Apache-2.0 means commercial use without asking permission, without a sales call, without an English contract your legal team takes three weeks to review. You can download the weights, fine-tune on your domain — your jargon, your code, your vertical — and run it wherever you want. It’s the difference between renting generic intelligence and building your own asset on an open foundation.

The shift almost nobody is watching: measuring by tokens, not just by score

Here’s the point I want you to take from this article, because it’s the one that changes how we choose models.

For two years, the conversation was a race of scores: who scores highest on a benchmark. Inkling proposes another axis. With its “controllable thinking effort” — you can modulate how much the model “thinks” — the relevant metric stops being only how accurate and becomes how much it costs you to be accurate. And there the numbers are concrete: according to Artificial Analysis, Inkling averages ~25K output tokens per Intelligence Index task, versus ~37K–43K from leading open-weights competitors like GLM-5.2, Kimi K2.6, or DeepSeek v4 Pro. That’s about 40% fewer tokens for comparable performance.

And it’s not sacrificing capability to achieve it. On Terminal Bench 2.1 — the benchmark for agentic terminal tasks — Inkling scores 63.8% and beats Nemotron 3 Ultra (56.4%). On other axes it holds the line: AIME 2026 at 97.1%, VoiceBench at 91.4%, MMMU Pro at 73.5%. In other words: it wins on the agentic task and uses fewer tokens.

Note for anyone familiar with the original brief: the circulating version said Inkling “matched Nemotron 3 Ultra on Terminal Bench using a third of the tokens”. The actual numbers are better in one part and different in the other — Inkling beats Nemotron there, and the token advantage (~40%) is measured over the Intelligence Index, not Terminal Bench. It’s worth getting it right: the correct number is stronger than the myth.

Why is this strategic and not a technical detail? Because in production, you pay by the token. A model that solves the same task with 40% less output is 40% cheaper to operate at scale, with less latency. For Latin American teams competing with tighter infrastructure budgets than Silicon Valley, token efficiency isn’t a footnote: it’s the line that separates a viable pilot from one that doesn’t close. Maximum score is paid for by whoever can afford it. Efficiency is leveraged by whoever knows how to read their bill.

What I’d do with this if I led a team in the region

Three concrete moves.

First, evaluate Inkling as a fine-tuning base, not as a direct replacement for your current API. The advantage doesn’t appear in the generic base model; it appears when you fine-tune it on your domain. Tinker lowers the barrier to entry (50% launch discount), but the deeper decision is to treat the model as your own infrastructure, not as a rented service.

Second, incorporate cost-per-task into your selection criteria from now on, whether you run Inkling or not. Stop comparing models only by their leaderboard number and start measuring tokens-per-task on your use cases. It’s a discipline that pays for itself.

Third, if data or latency matter, look at self-hosting with fresh eyes. Apache-2.0 plus an NVFP4 checkpoint make running your own model — or the Inkling-Small preview for cost-sensitive loads — stop being a weekend experiment and start being a serious architecture option. With the honest caveat: self-hosting documentation is still thin, so this is terrain for piloting, not for promising production next week.

The bottom line

Inkling won’t top the leaderboards, and its team knows it and said so. That frankness is precisely the signal. The question it raises isn’t “which is the smartest model?”, but “whose intelligence are you building with, and how much does each response cost you?”. For Latin America, where the edge rarely comes from compute budget, starting to answer those two questions — model ownership, cost per token — can be more decisive than chasing the last benchmark point.

Murati bet that the future is open, customizable, and measured in efficiency. It’s worth taking that bet seriously.

Sources:* Inkling: Our open-weights model — Thinking Machines Lab (15 jul 2026)