Thinking Machines Inkling: A 975B Open-Weights Model

Thinking Machines Lab has released Inkling, its first open-weights foundation model, and the architecture is worth studying closely. Inkling is a mixture-of-experts transformer with 975 billion total parameters but only 41 billion active per token — a sparsity ratio that lets a frontier-scale model run at a fraction of the inference cost a dense model of equivalent size would demand. The weights are available free on Hugging Face, and that combination of scale and openness is what makes this release notable.

• 975B total parameters, 41B active per token — a mixture-of-experts design pretrained on 45 trillion tokens of text, images, audio, and video
• 1M-token context window with natively multimodal reasoning (text-only output)
• Inkling-Small preview at 276B total / 12B active targets lower-cost deployments
• Free weights on Hugging Face; API access via TogetherAI, Fireworks, Modal, Databricks, and Baseten

What Makes Inkling's Architecture Different?

The headline number is 975B, but the number that matters for practitioners is 41B. In a mixture-of-experts model, a routing layer selects a small subset of specialized sub-networks for each token, so the compute cost per token tracks the *active* parameter count rather than the total. Inkling's roughly 24:1 ratio places it firmly in the sparse-frontier camp that has come to define open-weight releases this year — a lineage our AI coverage has followed through releases like Kimi K2.7's open-weight coding model.

Pretraining spanned 45 trillion tokens across text, images, audio, and video, giving the model native multimodal reasoning. Output remains text-only. The 1M-token context window puts whole codebases and document collections within a single prompt.

One feature deserves attention: controllable thinking effort, which lets a caller trade inference cost against latency at request time. Rather than choosing between a fast model and a careful one — the tiering approach we examined in our breakdown of the GPT-5.6 model tiers — Inkling exposes that dial inside a single model.

How Efficient Is Inkling in Practice?

The model was trained on NVIDIA GB300 NVL72 systems. TechCrunch reports that Inkling reaches comparable coding performance using roughly a third as many tokens as NVIDIA's Nemotron 3 Ultra — a token-efficiency claim we'd flag as coming from a single outlet and worth independent benchmarking.

A preview variant, Inkling-Small, runs 276B total parameters with 12B active, aimed squarely at teams that want the architecture without the deployment footprint. Fine-tuning is available through the company's Tinker platform, with 64K and 256K context options.

Why the Candor Matters

The most unusual thing in the announcement is a sentence most labs would have cut. Thinking Machines states plainly that Inkling is "not the strongest overall model available today, open or closed."

That candor is a genuine contribution. A well-funded lab publishing frontier-scale weights while explicitly declining to overclaim sets a norm the field benefits from. Open weights at this scale expand what independent researchers, small teams, and academic labs can build on — and an honest capability statement lets them choose the right tool instead of discovering the gap in production.

For developers, the practical takeaway is straightforward: a 1M-context, natively multimodal, 41B-active model is now downloadable, inspectable, and fine-tunable. That is a meaningful addition to the open-weight toolkit, whatever its position on a leaderboard.

Sources: Thinking Machines Lab — July 15, 2026; TechCrunch — July 15, 2026; Hugging Face — July 15, 2026.