Murati's Thinking Machines enters open-weights race with efficiency pitch, not top spot

Thinking Machines Lab, the start-up Mira Murati founded after leaving OpenAI's chief technology officer role and raising $2 billion in seed funding, has put its first open-weights model into the market — and it is not claiming the crown. Inkling, a 975-billion-parameter mixture-of-experts model with 41 billion active parameters per token, is positioned instead as infrastructure: a base model built for enterprises to fine-tune rather than run off the shelf.

The architecture is built for cost control. Inkling routes each token through 6 of 256 experts per layer, mixes sliding-window and global attention in a 5:1 ratio, and was trained on 45 trillion tokens spanning text, images, audio and video. The payoff, according to the lab, is reasoning performance — AIME 2026 at 97.1%, GPQA Diamond at 87.2%, SWE-bench Verified at 77.6% — delivered at around a third of the token cost competitors charge for comparable scores. Terminal Bench 2.1 came in at 63.8% and FORTRESS adversarial safety at 78.0%. A smaller sibling, Inkling-Small, runs 12 billion active parameters against 276 billion total, with full weights to follow once testing wraps.

Distribution is already wide. Weights sit on Hugging Face, and the model is live through Thinking Machines' own announcement as well as APIs from Together AI, Fireworks, Modal, Databricks and Baseten. The lab's Tinker fine-tuning platform carries a 50% discount for a limited time, an incentive clearly aimed at pulling developers into the ecosystem early rather than waiting for them to shop around.

The company is candid about where Inkling sits in the field. "Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization," the lab said in its launch post. That is a deliberate retreat from the leaderboard fight and a bet that cost-per-token and fine-tuning flexibility matter more to enterprise buyers than topping a benchmark table.

Inkling arrives the same week as Moonshot's 2.8-trillion-parameter Kimi K3, underscoring how crowded the open-weights tier has become, with both US and Chinese labs racing to narrow the gap with closed frontier systems.

Thinking Machines is not trying to beat GPT- or Gemini-class models head-on; it is trying to become the base layer other companies build on, monetised through Tinker and discounted API access rather than raw model sales. That is a distribution play as much as a research one, and the five-platform rollout at launch suggests the lab prioritised reach over exclusivity.

The risk is that "good enough, cheaper" is a crowded pitch this week alone — Kimi K3's arrival shows rivals are scaling total parameters even faster. Thinking Machines' edge, if it holds, will show up in fine-tuning uptake on Tinker rather than in benchmark charts, and that is the metric worth watching next.