Thinking Machines Lab released Inkling, its first open model: a multimodal Mixture-of-Experts with 975B total parameters / 41B active, with 256 routed experts (6 active per token), 1M context, and training on 45 trillion tokens of text, image, audio, and video. The weights are open-weights on Hugging Face, it runs on the Tinker platform, and it includes an adjustable 'thinking effort' to balance quality and cost per token.
For years, the phrase "a genuinely powerful open model" came with fine print: either they were tiny models lined up well behind the closed ones, or the weights were hidden behind an API with a waitlist. That's why July 15 raised so many eyebrows when Thinking Machines Lab — the research lab founded by Mira Murati, former OpenAI CTO, with one of the most expensive technical teams on the planet — released Inkling, its first model, and dropped the full weights on Hugging Face as open-weights, while also running it on its Tinker platform and on partner APIs. Not a demo with a pretty screenshot: the weights, to download and run on your own infrastructure. When a lab of that caliber chooses to lead with open, the gesture itself is the story.
## A near-trillion parameters that fit your budget
Inkling is a multimodal Mixture-of-Experts (MoE) with 975 billion total parameters, but — and here's the MoE magic — it activates only 41 billion for each token it processes. The useful analogy: every layer has 256 routed specialists on the payroll (plus a couple always on call), but for each token it wakes only the six it needs; you pay for six units of compute, not two hundred fifty-six. Add context of up to 1 million tokens, training on 45 trillion tokens of text, images, audio, and video, and a lighter preview variant — Inkling-Small, at 276B total / 12B active — for anyone who doesn't need the flagship. It's the kind of architecture that makes a huge model, paradoxically, viable to operate.
## The lever almost nobody is watching: adjustable effort
The most interesting thing for builders isn't the parameters, it's a subtler detail: Inkling lets you control how much it "thinks" before answering, to balance performance against token consumption. The creators themselves illustrate it with a concrete figure: it matches Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly a third of the tokens. Translated into your bill: every extra token the model spends reasoning is money leaving your account, and here you can decide per request whether the task deserves deep mode or a quick answer will do. It's a quality-versus-cost dial placed in your hands. And the lab doesn't hide the honest caveat: it says plainly that Inkling "is not the strongest model available today, open or closed," but rather a solid open base to customize and fine-tune. That candor, in a market addicted to claiming superiority on every benchmark, is refreshing.
## Why 'open-weights' changes the starting line
The fact that the weights are available to download is the difference that becomes the headline. Combine that with an MoE that activates "only" 41B parameters per token and checkpoints prepared for modern hardware — including an NVFP4 version for efficient inference on NVIDIA Blackwell systems — and suddenly a small team can consider running a near-trillion model on its own infrastructure, or fine-tuning it on its domain via Tinker, without being tied to a closed API. It doesn't replace frontier models for everything — the authors themselves admit as much — but it does redefine the starting line for what a modest team can have under its control and its budget. It's worth reviewing the exact usage terms on the model card on Hugging Face before taking it into commercial production, though.
## What it means for those building with AI
The deeper lesson of Inkling is that competitive advantage is shifting from "which model do I have" to "how do I govern it." An adjustable thinking effort isn't a benchmark trick: it's exactly the kind of lever a product's profitability depends on — whether it makes money or eats its margins token by token. At NeuralOS we think from that quiet arithmetic — which is why cost guardrails (per-user budgets, per-workspace ceilings, streaming to abort early) are part of the design and not an optional extra. And by conviction we're model-agnostic: an open one like Inkling enters the same conversation as a closed frontier model, because you pick the engine by task and don't marry any of them. No smoke: we're not promising to serve Inkling out of the box today. What is part of the design is the discipline of deciding how much each request thinks — and how much it costs.