Inkling 975B

Thinking Machines' first model: an open-weights Mixture-of-Experts transformer with 975B total and 41B active parameters, pretrained on 45 trillion tokens of text, images, audio and video. It reasons natively over all three input modalities through an encoder-free architecture that embeds image patches and audio spectrograms alongside text tokens, so it can transcribe speech, follow spoken instructions and reason over recordings without a separate pipeline. Thinking effort is continuously controllable rather than stepped, and the lab's case for it is token efficiency: reaching a given score with fewer tokens than comparable models. Thinking Machines is direct that this is a base to customise rather than a frontier model, saying it is not the strongest model available today, open or closed.

Inkling is the first model Thinking Machines has trained itself, and the post announcing it spends most of its length explaining a choice rather than a score. The lab did not set out to win a leaderboard. It set out to publish a base that other people can take apart: full weights, a broad and deliberately balanced capability profile, and a fine-tuning path that exists on day one rather than as a promise. The argument behind that is a claim about where the remaining value sits - that many real problems are not solved well by even the best generalist model, and that the gap is closed by fine-tuning on knowledge a lab does not have and cannot buy.

The architecture follows from it. Inkling is a Mixture-of-Experts transformer of 975B total parameters with 41B active, 256 routed experts per layer and six active per token, using sigmoid routing, relative positional embeddings rather than rotary ones, and alternating sliding-window and global attention. It was pretrained on 45 trillion tokens spanning text, images, audio and video, and it carries a context window of up to a million tokens. None of that is arranged around a single headline capability; it is arranged around being a reasonable starting point for a lot of different endings.

Multimodality here is architectural rather than adjacent. Thinking Machines describes an encoder-free design: images arrive as 40x40 pixel patches through a four-layer hMLP, audio as dMel spectrograms through a light embedding layer, and both are then processed jointly with text tokens in the same stack. The practical consequence is that audio is a first-class input, not a transcription step bolted to the front - the model can transcribe speech, follow spoken instructions, answer questions about a recording and reason over longer-form audio without a separate pipeline in between.

The efficiency claim is the one the lab makes hardest, and it is about tokens rather than latency. Thinking effort on Inkling is continuous rather than stepped, set as a value between 0.2 and 0.99, and the lab's charts are drawn as score against tokens spent instead of score alone. On that framing it reports Inkling reaching a given score at fewer tokens than the models it compares against, matching Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens. Its reported figures include 77.6% on SWEBench Verified, 63.8% on Terminal Bench 2.1, 87.2% on GPQA Diamond, 97.1% on AIME 2026, and on the multimodal side 73.5% on MMMU Pro and 91.4% on VoiceBench.

Thinking Machines is unusually direct about the ceiling. Its own words are that Inkling "is not the strongest overall model available today, open or closed", and that this is a first release in a family it intends to keep building on - it expects the multimodal side in particular to improve. Read that as the shape of the offer: reach for Inkling when you want a capable, cheap, genuinely multimodal model you can inspect, host or specialise, and reach for something else when you want the highest possible score on a task you are not going to fine-tune for.