Thinking Machines Lab Releases Inkling: An Open-Weights Foundation Model Built for Customization, Not Leaderboard Dominance

Mira Murati’s new lab ships a 975B-parameter mixture-of-experts model with full open weights and a fine-tuning platform — and says outright it is not the strongest model available. That framing is the strategy.

What Happened

Thinking Machines Lab — founded by Mira Murati, who served as CTO of OpenAI until late 2024 — released its first model on July 15, 2026. Inkling is a mixture-of-experts foundation model carrying 975 billion total parameters, with 41 billion active per forward pass. It was pretrained on 45 trillion tokens spanning text, images, audio, and video, supports a one-million-token context window, and ships alongside Inkling-Small, currently in preview. Full weights are available on Hugging Face; inference is live across Together AI, Fireworks, Modal, Databricks, and Baseten.

Thinking Machines is explicit in its announcement that Inkling is “not the strongest overall model available today.” The lab frames it instead as a broad, balanced base model designed to be customized — the flagship model for Tinker, its fine-tuning platform. Key features include multimodal reasoning, agentic coding and tool use, calibrated forecasting, and a controllable “thinking effort” dial that lets users trade inference token cost against performance depth. The lab describes Inkling as “just the start… our first release in a model family we will continue to build on,” anchored to a stated mission of building AI that “extends human will and judgment.”

Benchmark results, by the lab’s own reporting, are respectable — strong performance on reasoning and coding suites — but self-described as not category-leading. That is a critical caveat: everything here flows from Thinking Machines’ own first-party announcement, including the benchmark scores and the “not the strongest” characterization. Independent third-party evaluations have not yet been published at time of writing.

The Structural Read

The correct frame for Inkling is not “how does it score on MMLU?” It is: what theory of AI value capture does this product embed, and is that theory likely to be right?

Read one: a deliberate bet that frontier capability commoditizes. The core wager is that as raw model intelligence becomes increasingly available — from open-weights releases, from efficiency gains compressing the cost curve, from the proliferation of capable base models — the durable competitive advantage shifts away from owning the single most powerful system and toward the ability to adapt a model to a specific domain, workflow, and proprietary dataset. If that is right, then not chasing the frontier is rational capital allocation: you are not losing a race you have chosen not to run. If the frontier-max camp is right — if the strongest closed model continues to compound advantages in users, revenue, and data flywheels — then an open, deliberately-not-strongest model is a structurally harder sell against incumbents (OpenAI, Anthropic, Google) that hold deep distribution, enterprise relationships, and the strongest closed systems today. That tension is unresolved, and Thinking Machines’ bet is a thesis, not a demonstrated win.

Read two: open weights plus a fine-tuning platform is a different value-capture model. Rather than monetize inference tokens on a proprietary closed model, Thinking Machines gives the weights away and sells the customization toolchain — Tinker — capturing the adaptation layer instead of the raw intelligence layer. This is a picks-and-shovels position: the value accrues not to whoever runs the most impressive general model, but to whoever owns the workflow by which organizations make a model their own. The signal that this lane has real economics is already visible in developer behavior: open models carry a large and rising share of developer inference tokens, a trend documented in open-weights usage data tracked via OpenRouter. The counterargument is that the fine-tuning and customization market is already competitive — Hugging Face, Replicate, major cloud providers, and the frontier labs themselves all offer fine-tuning surfaces — and that owning the adaptation layer is only defensible if the base model earns developer loyalty first.

Read three: Inkling sharpens the neolab thesis. Thinking Machines joins a pattern of well-credentialed teams choosing non-consensus lanes rather than competing directly with frontier incumbents. Nous Research went open and decentralized; Sutton’s Oak Lab bet on continual learning from experience. Thinking Machines bets on customization infrastructure and an open base. All three lanes share a common structural argument: the long-run moat is not renting the best available general model, but owning a fine-tuned, data-adapted system built on a model you control. That argument — most clearly articulated in Nadella’s framing of the enterprise learning loop — says the enduring advantage accrues to whoever closes the loop between proprietary data and a model that can be continuously shaped by it. Inkling plus Tinker is a direct attempt to be the infrastructure on which that loop runs.

Three Implications