On July 15, Thinking Machines Lab announced Inkling, the first model the company has trained from scratch and shipped with full weights available. It is the release the AI world had been waiting on since Mira Murati left her post as OpenAI’s CTO in September 2024 to start the lab — and now that the initial coverage cycle has passed, the more interesting question is not whether Inkling tops leaderboards, but what kind of open-weights model it is trying to be. The answer is unusual: Thinking Machines says so itself, right in the announcement, that “Inkling is not the strongest overall model available today, open or closed.”
That is not a sentence frontier labs normally publish on launch day. It is the key to understanding the entire release.
A trillion parameters, 41 billion at a time
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, of which 41 billion are active on any given forward pass. That sparse design is what makes a nearly-trillion-parameter model practical to serve: you pay inference costs closer to a mid-sized dense model while keeping the capacity of a much larger one. The model supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio and video.
The multimodality is native, not bolted on. Inkling reasons directly over text, images and audio, and exposes what Thinking Machines calls controllable thinking effort — the amount of reasoning the model spends on a task can be dialed up or down, including from inside an agent harness. For teams building products where latency and cost budgets vary wildly between features, that knob matters more than a benchmark point or two.
Inkling is also explicitly the first in a family. Alongside the flagship, the lab shared a preview of Inkling-Small, a lighter model with 12 billion active parameters trained on a similar recipe, aimed at workloads where cost and latency dominate.
The pitch: a base model, not a champion
Most open-weights launches lead with evaluation tables. Thinking Machines leads with a different claim: that a combination of qualities — broad multimodal capability, efficient and controllable reasoning, and day-one availability for fine-tuning — makes Inkling a good base for customization, even though stronger overall models exist. The announcement frames the model as “broad, balanced” rather than narrowly optimized, trained across agentic, reasoning, coding, instruction-following, factuality, vision and audio tasks. Its own benchmark chart plots Inkling against Nemotron 3 Ultra, GLM 5.2, GPT 5.6 Sol and Claude Fable 5 — and the point of the chart is breadth of coverage, not a sweep of wins.
This positioning is coherent with the company’s stated mission of building AI that “extends human will and judgment,” and with everything it shipped before Inkling: the Tinker platform for customizing models, a previewed collaborative AI system, and published research. The lab’s bet is that the durable value in open weights is not raw capability — which decays against next month’s release — but adaptability. A model you can reshape to your own data, tools and constraints stays useful after the leaderboard moves on.
For agentic work specifically, Thinking Machines says it trained Inkling to run inside a variety of coding and agent harnesses and randomized the tool set and schema during training, so the model is not brittle about any particular tool-calling format. It runs inside the OpenCode harness in the lab’s own demos. Independent early coverage points the same direction: Decrypt’s testing called the model’s MCP tool-use performance “genuinely impressive” while noting the price-to-performance picture is more complicated — a fair summary of a model engineered to be a dependable agent substrate rather than a record-setter.
The model that fine-tuned itself
The launch demo is the thesis stated as theater. Thinking Machines asked Inkling to fine-tune itself using Tinker toward a target no system prompt can reliably enforce: a lipogram model that never uses the letter “e.”
Inkling wrote its own objective file and scoring function, generated synthetic training data, launched the fine-tuning job through the Tinker API, evaluated the result against its own base weights, and then staged a self-update that relaunched the harness on the new checkpoint. The whole loop completed in about 27 minutes in the recorded run, ending with the updated model answering questions in fluent e-free English. The demo is playful, but the mechanism it demonstrates is the product: specify a behavior, and the model plus Tinker close the loop from intent to new weights without a research team in between.
That is also why the release ships with the Inkling Playground in the Tinker console — a developer-facing chat interface, because Thinking Machines argues picking a base model to fine-tune is partly a qualitative judgment that “comes from playing with it.” Benchmarks tell you whether a model can do something; only conversation tells you whether it does it in a way your product can live with, and for a model whose entire pitch is customization, that hands-on feel is part of the evaluation, not a luxury.
Where you can actually run it
Access on day one is broader than most from-scratch releases manage. The full weights are on the model card on Hugging Face, with a pre-quantized Inkling-NVFP4 variant published alongside for efficient deployment on supported hardware. Hosted inference is live on OpenRouter, which means any application already speaking the OpenRouter API can trial Inkling with a one-line model-string change. And fine-tuning is available immediately on Tinker, which is the distribution channel the whole release is really about.
For self-hosters, the usual sparse-model math applies: 975B total parameters is a serious storage and memory footprint even though only 41B activate per token, so the NVFP4 quantization and the forthcoming smaller family members are the realistic paths for most local deployments. Inkling-Small’s 12B active parameters put it much closer to commodity-hardware territory.
What this signals for open weights
Inkling lands in the middle of an active argument about what “open” AI should mean — a debate over licensing definitions and, in Washington, over whether open-weight releases should face policy restrictions at all. Against that backdrop, a well-funded American lab founded by one of the most prominent figures from OpenAI choosing full-weights release for its very first model is a meaningful data point, whatever one thinks of any single benchmark.
It also sharpens the competitive question for the open ecosystem. The strongest open-weights models of the past year have overwhelmingly come from outside the United States, and the standard critique of Western labs has been that they keep their best work closed. Thinking Machines’ answer is not to out-benchmark anyone. It is to make customization — the thing enterprises actually do with open weights — the headline feature, with a from-scratch multimodal base, a controllable reasoning budget, and a first-party fine-tuning platform wired in from day one.
Whether that bet pays off depends on what gets built on top of it. A base model’s success is measured in derivatives, not launch-week discourse — and by that metric, the scoreboard for Inkling starts now, in the fine-tunes that show up on Hugging Face over the next few months. The model card, the weights and the tooling are all public. The rest is up to the people Murati’s lab says it built this for.
