Inkling Is a 975B Open-Weights Model Under Apache 2.0

A Big Open-Weights Release With an Unusually Honest Pitch

Thinking Machines Lab released Inkling on July 15, 2026 — its first open-weights model, and by parameter count one of the largest ever published under a permissive license. The weights are on Hugging Face under Apache 2.0. What makes the release worth reading twice is not the size, though: it is the framing. The lab states plainly that Inkling "is not the strongest overall model available today," and positions it as a broad, adaptable base for organizations to fine-tune rather than a leaderboard entry.

• Mixture-of-experts transformer with 975B total parameters and 41B active; an Inkling-Small variant runs 12B active
• Pretrained on 45 trillion tokens across text, images, audio, and video, with native reasoning in all four modalities
• 1M-token context window, released on Hugging Face under Apache 2.0
• Post-training used large-scale reinforcement learning with more than 30M rollouts

What Does "975B Total, 41B Active" Actually Mean?

Inkling is a mixture-of-experts model, so only a fraction of its parameters fire on any given token. The full network holds 975 billion parameters, but roughly 41 billion are active per forward pass — and the smaller Inkling-Small variant drops that to about 12 billion. In practice this is the architecture that lets a very large model stay servable: you pay for the capacity in memory, not in per-token compute.

That design has become the default at the top of the open-weight field, and it is the same broad approach behind the largest open-weight model released this month. The interesting variable is no longer whether a lab uses MoE, but what the active-parameter budget is and how the routing was trained.

Four Modalities, Trained Together

The pretraining mix spans text, images, audio, and video across 45 trillion tokens, and Thinking Machines describes reasoning as native across all four rather than bolted on through adapters. Post-training leaned heavily on reinforcement learning — over 30 million rollouts — with the lab reporting that reasoning performance improved log-linearly through the run. That last detail is the one researchers will chase: a clean scaling relationship in RL post-training is a much more useful thing to know than a single benchmark figure.

How Can You Actually Run It?

The weights are downloadable under Apache 2.0, which is about as unrestricted as a major-lab release gets — commercial use, modification, and redistribution are all permitted. For teams that would rather not host a near-trillion-parameter model themselves, Inkling is served through Together AI, Fireworks, Modal, and Databricks. Thinking Machines is also running a limited-time 50% discount on its Tinker fine-tuning platform, which is where the "adaptable base" pitch is meant to land.

Artificial Analysis rates Inkling the leading US open-weights model, and independent evaluation of a model this fresh is worth watching over the next few weeks. Inference-serving capacity has been expanding quickly on that front too — Fireworks AI's recent raise was aimed squarely at making models this size practical to run.

The Part Worth Copying

Plenty of releases oversell. Shipping a very large model with permissive weights and an explicit acknowledgment that it is not the strongest option available is a rarer move, and a healthier one — it tells developers what the model is for instead of what it might win. Our full AI coverage has more on how the open-weight field is evolving.

Sources: Thinking Machines Lab — July 15, 2026; TechCrunch — July 15, 2026; Artificial Analysis — July 2026.