Thinking Machines Lab releases its first model, an open-weights system called Inkling

Thinking Machines Lab released Inkling on July 15, its first model and the first one it has trained from scratch. The full weights are on Hugging Face. What is unusual is not the release. It is the sentence the company put in its own announcement.

Inkling is not the strongest overall model available today, open or closed.

Labs do not usually say that about the thing they just shipped. Thinking Machines said it in the fourth paragraph, and the rest of the post is an argument for why that is beside the point.

What the thing actually is

Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active on any given token. It takes a context window up to 1 million tokens and was pretrained on 45 trillion tokens of text, images, audio, and video. Alongside it, the lab previewed Inkling-Small, a 276-billion-parameter model with 12 billion active, and said it will release those weights once testing is done.

The model reasons natively over text, images, and audio instead of bolting a vision model onto a text model. The audio and vision paths are encoder-free: audio goes in as dMel spectrograms, images as 40-by-40-pixel patches run through a light four-layer network, both fed into the same stream as text tokens. That is a design choice aimed at real-time voice and vision use, which is where the lab has said it is heading with its interaction-model work.

The pitch is the fine-tuning, not the model

Read past the specs and the strategy is plain. Inkling is a base model you are meant to customize, and the product it feeds is Tinker, Thinking Machines' fine-tuning platform. The model is available to fine-tune there today, at a 50 percent discount for now, with a browser playground to try it before you commit compute. The open weights are what make that pitch land: download it, adapt it, run it where you want, or rent it through APIs on Together, Fireworks, Modal, Databricks, and Baseten.

So "not the strongest model" reads less like humility and more like positioning. The frontier-benchmark race is not the game Inkling is playing. The bet is that enough customers want a customizable open base more than they want the highest benchmark score. This release does not tell you whether that bet is right.

The one number worth pulling out

The feature that earns attention is controllable thinking effort. You set how hard the model works, from a light pass to a long one, and it spends tokens accordingly. Thinking Machines reports that Inkling matches Nemotron 3 Ultra on the Terminal Bench 2.1 coding benchmark using roughly a third of the tokens. Tokens are cost and latency, and for anything you run millions of times, a third of the tokens is a real number, not a slide.

The mechanism is worth a line because it is a little strange. The lab did not hand-tune an effort dial. During reinforcement learning it varied a per-token cost across training samples, and the model learned on its own to spend more or fewer tokens depending on the instruction. They also noticed the chain-of-thought getting terser as training went on, dropping articles and connectives while still reaching the same answer, an efficiency the reward did not ask for directly.

Where it sits on the benchmarks

Give the lab credit for how it reported the numbers. Everything is run at maximum effort and stated as such, and where possible it uses third-party scores from Artificial Analysis instead of its own runs. On Terminal Bench it flagged that some solutions were contaminated by web search and zeroed them. That is more traceable sourcing than most launch posts bother with.

The scores themselves put Inkling behind the frontier closed models and behind some open ones. On Humanity's Last Exam, text only, it scores 29.7 percent, against 40.1 for GLM 5.2 and 53.3 for Claude Fable 5. On SWEBench Verified it lands at 77.6 percent, in the same band as Kimi K2.5 and below Kimi K2.6 and GPT 5.6 Sol. It clears AIME 2026 at 97.1 percent, where everything at this tier is near the ceiling. The picture matches the company's own framing: broad and balanced, ahead of nothing in particular.

Uncensored by design, and they measured it

Thinking Machines trained Inkling for what it calls epistemics: calibration, instruction following, and "resistance to censorship." The last one is a stated training goal. The lab had Cognition run the model on its Propaganda and Censorship Eval, and reports Inkling shows strong patterns of refusing to censor.

Keep that separate from safety, because the two get confused. On FORTRESS, which tests whether a model refuses genuinely harmful requests about weapons and violence, Inkling posts 78 percent on the adversarial split, the best of the open-weights models the lab compared, and 98.6 on StrongREJECT. By its own measures, "will not censor political questions" and "will still refuse to help you build a weapon" are separate settings, and it tuned them separately. Those are the lab's own benchmarks, so read them as a claim to verify, not a settled fact.

The parts a builder will notice

The architecture leans on DeepSeek-V3's recipe, 256 routed experts plus 2 shared, 6 active per token, with a few departures. Thinking Machines swapped the now-standard rotary position embeddings for relative ones, saying they extrapolate better to long sequences, and interleaves local and global attention at a 5-to-1 ratio. Training used a hybrid optimizer, Muon for the big weight matrices and Adam for the rest, on NVIDIA's GB300 NVL72 systems, and scaled RL past 30 million rollouts.

One detail stands out. To bootstrap post-training, the lab ran an initial supervised pass on synthetic data generated by open-weights models including Kimi K2.5, Moonshot's Chinese release. A US lab founded by OpenAI's former CTO seeding its first model on outputs from a Chinese open model is a small sign of how tangled the open-weights supply chain has become.

Inkling is a debut, and the lab says so plainly: the first of a family, with more to come. The weights are public, the fine-tuning is where the money is meant to be, and the benchmarks are candid about landing mid-pack. For a first model, those are coherent choices. Whether they add up to a business is the question this release does not try to answer.