Not the strongest model — but possibly the most useful open model.
On July 16, Thinking Machines Lab (TML) released its first foundation model, Inkling: 975 billion parameters, MoE architecture, Apache 2.0 open weights. In today’s market those numbers barely qualify as news.
The real news is the first sentence of their official blog: “Inkling is not the strongest model available today — open or closed.”
A company founded by former OpenAI CTO Mira Murati and former applied-research VP Luna Wang, backed by a $2 billion seed round after training on 45 trillion tokens, opens with a concession? Of course not. What they are saying is: the dimension on which open models compete is undergoing an irreversible switch.
For two years the open-source narrative has been driven by one question — “can it catch GPT/Claude?” From Llama to DeepSeek to GLM, every release measured itself against the closed flagship: the high-performance race, where whoever sits nearest the closed ceiling on the leaderboards wins. Inkling replaces the question at the root.
Three Strange Things Inkling Did
TML’s team came from OpenAI’s core; they lack neither talent nor capital, and could have played the strongest-open-model script. They didn’t. Three deliberately counterintuitive decisions:
1. Publicly declaring “not the strongest”
Review the launch notes of every mainstream open model in recent years — none volunteered “we are not the strongest.” Inkling is the first. That is not modesty; it is a clear-eyed definition of the target market. TML spent the past year serving enterprises through the Tinker fine-tuning platform, and they learned what any enterprise buyer knows: customers choosing models do not care about the gaps inside the leaderboard’s top ten.
2. Deliberately training the model to sometimes think less
Inkling ran large-scale asynchronous RL in post-training — over 30 million rollouts — but the objective was not making the model smarter. It was teaching it to read the table and order accordingly. TML’s mechanism is elegant: by modifying system messages and assigning different per-token costs to different tasks, the model learned to judge which problems deserve deep pursuit, which tasks can stop early, and whether another paragraph of reasoning buys enough performance to justify itself. The result: on Terminal Bench 2.1, Inkling matches NVIDIA’s Nemotron 3 Ultra while generating one-third the tokens. In engineering terms: this is not teaching a model to think deeper — it is teaching it to think more profitably.
3. Letting the model fine-tune itself
TML staged a audacious demo: Inkling writing its own fine-tuning jobs through Tinker, running the training, evaluating the results. The model became infrastructure for shaping itself. The demo looks like a stunt but encodes TML’s entire business logic — plasticity is the product. A model’s final capability boundary depends not on its pretraining ceiling but on whether enterprises can keep reshaping it with their own data.
The Three Axes of Usability — a New Evaluation Frame
Put the three decisions side by side and one logic emerges: once absolute capability passes a threshold, the enterprise standard for “good model” inverts. Call it the three axes of usability:
Axis one: token economics. How much task success does each token buy? Not a static metric — a function of reasoning intensity. A model that only thinks at full power and a model that compresses reasoning on easy tasks while escalating on hard ones are two entirely different products for an enterprise.
Axis two: completeness of the customization loop. Open weights plus a fine-tuning platform plus data-rendering tooling plus deployment adapters — all four, no exceptions. Weights without tooling are an engine without a manual. TML’s tml-renderer, shipped alongside, closes the most-neglected hole in that chain: sampling and fine-tuning format conversion for multimodal data.
Axis three: depth of native multimodality. Text, image, and audio entering joint learning from pretraining — not bolt-on vision or audio encoders added later. The bolt-on philosophy says “build a language model, then add senses”; the native philosophy says “interaction is multimodal by default, so the model should be.” Inkling positions native multimodality as the backend reasoning engine inside a collaborative model architecture, working asynchronously with a front-end real-time interaction model — a deeper design choice than “supports image input.”
Running Other Models Through the Three Axes
The frame is not Inkling-specific — it suddenly explains last year’s “strange” market phenomena. Why did DeepSeek V3 win global developer goodwill? Not by beating GPT-4o, but by pulling inference cost down an order of magnitude — it wins on axis one. Why is GLM 5.2 penetrating Chinese enterprise so fast? Not just capability — Zhipu ships a complete toolchain from fine-tuning platform to deployment. It wins on axis two. Why did finance and healthcare shift from “can open models be used” to “which one should we pick”? Because those industries’ real questions were never a percentage point on a leaderboard: can it fine-tune on my data? can I carry the inference bill? can it handle my mixed voice, image, and text input?
Ask those three questions of the top five open models and very few pass unanimously. Inkling does — not by coincidence. TML served enterprises through Tinker and modeled those needs directly into the design objectives.
What It Means
Inkling’s release closes an era and opens one. The old era: open-model competition was a performance contest — whoever approached the closed ceiling represented open source’s progress. That narrative drove countless “X model catches GPT” headlines while narrowing open source’s market position to “free near-substitute.” The new era: open-model competition is a usability contest. Once capability crosses the threshold, the moat moves from the capability ceiling to the cost of using capability well. Token economics, customization depth, multimodal completeness — together they define what a model “can become in an enterprise’s hands,” not “where it ranks on a board.”
That is the frame switch, and it explains why TML could say “we are not the strongest” on launch day — because “strongest” was never the contest they entered.
What To Do With This
-
If you run technology selection: next time you evaluate an open model, look past HLE and SWE-bench. Ask: how many tokens does it need to finish my tasks? How complete is its fine-tuning toolchain? Is its multimodality native or attached? If you cannot answer, no score will save your production environment.
-
If you build models: Inkling’s RL-learned thrift deserves deep study. In cost-sensitive commercial settings, a model that throttles itself beats a model that only outputs at full power — a direction almost nobody has engineered yet. Opportunity.
-
If you invest: the three axes make a new diligence frame for model companies. A startup competing only on “strongest” will exhaust itself; differentiated positions on all three axes carry more commercial value than any leaderboard crown.
Sources: Thinking Machines Lab blog — Introducing Inkling · 36Kr — “Inkling, America’s non-chasing open-source king”
