Thinking Machines Lab's first model — a large, natively multimodal open-weights LLM built to be customized, not rented.
Last verified · 2026-07-17 · by Moe Ameen
Inkling is the first in-house model from Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. It was released on July 15, 2026 as open weights under the permissive Apache 2.0 license — an unusually open move for a frontier-scale lab. It is a natively multimodal mixture-of-experts model: 975 billion total parameters, but only about 41 billion active for any given token, so it runs far cheaper than its headline size suggests. It reads text, images, and audio and responds in text, with up to a 1M-token context window on the open weights (256K through the company's Tinker fine-tuning API). Thinking Machines says it was trained on roughly 45 trillion tokens of text, images, audio, and video.
The framing behind Inkling is a bet against one-size-fits-all AI: rather than a single closed general-purpose system you rent, it is an open base meant to be fine-tuned and specialized for your own use. Two features stand out for that. Controllable "thinking effort" lets you dial reasoning up or down to trade speed and cost against depth, and the model is built to give calibrated answers — flagging uncertainty rather than guessing. Independent benchmarking by Artificial Analysis placed it as the leading U.S. open-weights release at launch, ahead of other open models on its Intelligence Index while using notably fewer output tokens per task.
You can download the weights from Hugging Face, fine-tune on the company's Tinker platform, or reach hosted inference through partners like Together AI, Fireworks, Modal, Databricks, and Baseten; it runs in SGLang, vLLM, llama.cpp, and the transformers library. A lighter preview, Inkling-Small (276B total, ~12B active), targets low-latency, low-cost workloads.
One thing to be clear about: Inkling is a text-output model. It reads images and audio, but it does not generate images, video, or audio. It writes, summarizes, reasons, and can turn a recording or a screenshot into structured text — it renders no media and publishes nothing. And "open weights" is not the same as fully open source: the checkpoint and license are public, but the full 45-trillion-token corpus is not. Like any LLM, its output can be wrong and needs checking before it ships.
Inkling's most useful trick for creators is on the input side. It reads audio and images natively, so you can hand it a raw recording — a podcast episode, a sales call, a voice memo — or a screenshot, and get back structured text: a summary, a set of hooks, a script, a caption pack. Because the weights are open under Apache 2.0 and fine-tunable through Tinker, a team can train it on its own archive so those drafts land in its actual voice. What Inkling cannot do is turn any of that into a post. It outputs text only — no images, no video, no captions burned onto a clip, no publishing.
That is the exact line where Kompozy takes over. Feed an Inkling-drafted script or an ingested-recording summary into Kompozy and it generates the media the model can't: HeyGen persona and avatar shorts with burned-in captions, face-locked Persona Photos and Persona Tweets, multi-slide Carousels and Quote Graphics rendered pixel-exact through HyperFrames, plus full blogs and newsletters. Then it publishes — fanning each piece across the nine supported platforms (Instagram, Facebook, TikTok, YouTube, LinkedIn, X, Pinterest, Threads) plus Mailchimp email and blog, on a schedule, on autopilot, with a per-post review pipeline. So a single recording Inkling read becomes a finished week of on-brand posts instead of a text file. Inkling turns raw multimodal source into words; Kompozy turns those words into media and ships them.
Inkling is the first open-weights model from Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released on July 15, 2026 under the Apache 2.0 license. It is a natively multimodal mixture-of-experts model — 975B total parameters with about 41B active per token — that reads text, images, and audio and responds in text, with up to a 1M-token context.
The weights are freely downloadable under Apache 2.0, which allows commercial use with no per-token fee to the model itself. It is best described as "open weights" rather than fully open source: the checkpoint and license are public, but the complete training dataset is not. Your real cost is the GPU/inference infrastructure to run it or a hosted provider's pricing.
No. Inkling reads images and audio as input but outputs text only — it does not render images, video, or audio. To turn its text into finished posts you pair it with a generation and publishing engine like Kompozy, which produces the media and publishes across platforms.
Its scale and design: a 975B/41B mixture-of-experts with native multimodal input, up to a 1M-token context, controllable "thinking effort" to trade speed for depth, and calibrated answers that flag uncertainty rather than guess. At launch, Artificial Analysis rated it the leading U.S. open-weights release on its Intelligence Index, and it is token-efficient relative to peers.
Inkling writes the copy but does not publish. Bring an Inkling-drafted script or caption set into Kompozy to generate persona video, carousels, quote cards, blogs, or newsletters in your brand voice, then schedule and publish the set across TikTok, Reels, Shorts, X, LinkedIn, and more from one queue.