Mira Murati's Thinking Machines Ships Inkling, a 975B-Parameter Model Under Apache 2.0

The lab's first in-house model is a sparse mixture-of-experts system trained on 45 trillion tokens and released with full weights — and the company openly concedes it is not the strongest model available, open or closed.

Jul 16, 2026Source: TechCrunch

Mira Murati's Thinking Machines Ships Inkling, a 975B-Parameter Model Under Apache 2.0

Key Points

1.Thinking Machines Lab released Inkling on July 15: a 66-layer sparse mixture-of-experts model with 975 billion total parameters and about 41 billion active per token, trained on 45 trillion tokens of text, image, audio and video across Nvidia GB300 NVL72 systems, with a context window up to 1 million tokens.
2.The weights ship under Apache 2.0 on Hugging Face alongside an NVFP4 quantized checkpoint for Blackwell, making it one of the largest permissively licensed models ever released; a smaller preview, Inkling-Small, has 276 billion total and 12 billion active parameters.
3.The company states plainly that 'Inkling is not the strongest overall model available today, open or closed,' and does not monetize it — revenue comes from Tinker, its fine-tuning platform, whose clients include Bridgewater Associates.

Eighteen months after Mira Murati left the OpenAI CTO job to start it, Thinking Machines Lab has shipped its first in-house model. Inkling, released July 15, arrives with something the frontier labs do not offer: the full weights, under Apache 2.0.

The architecture

Inkling is a 66-layer decoder-only transformer on a sparse mixture-of-experts backbone — 256 routed experts plus two shared experts per layer, with six routed experts active per token. That yields 975 billion total parameters but only about 41 billion active per token. It was trained on 45 trillion tokens spanning text, image, audio and video, on Nvidia GB300 NVL72 systems, and handles context windows up to 1 million tokens. The model card lists inputs as text, image and audio, with text-only output — video is in the training mix, not the input surface.

How to run it

Weights are on Hugging Face, with an NVFP4 quantized checkpoint for Blackwell hardware. Running it is not casual: BF16 needs roughly 2TB of VRAM, NVFP4 about 600GB. It is supported in Transformers, SGLang, vLLM, llama.cpp and Unsloth, served via TogetherAI, Fireworks, Modal, Databricks and Baseten, and available for fine-tuning on the company's own Tinker platform. A preview of Inkling-Small — 276 billion total, 12 billion active — uses the same recipe.

An unusually honest launch

Company-reported benchmarks put Inkling at 97.1% on AIME 2026, 87.2% on GPQA Diamond, 77.6% on SWE-bench Verified and 46.0% on Humanity's Last Exam with tools; none are independently verified. The lab claims Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra for equivalent coding performance. But it declines the usual claim: "Inkling is not the strongest overall model available today, open or closed," its blog states.

The business behind it

Founded in February 2025, Thinking Machines raised a $2 billion seed at a $12 billion valuation in July 2025 — the largest seed in AI history — led by Andreessen Horowitz with Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street. A follow-on reportedly sought at $50 billion collapsed by January 2026 without a deal, with backers wanting a stronger product record. Inkling is that record. It is not monetized; revenue comes from Tinker, whose clients include Bridgewater Associates.

Why It Matters

A near-trillion-parameter model under Apache 2.0 is the most permissive release at this scale to date, and it lands as US labs face pressure over China's lead in open weights. The commercial logic is the interesting part: give away the model, sell the fine-tuning platform — a bet that in a world of abundant weights, the durable margin sits in adaptation rather than inference. Murati's investors walked away from a $50 billion round in January for want of shipped product; the terms of any next raise will show whether this changes their answer.

Filed underAI Models Open Source AI #Inkling #Mira Murati #Thinking Machines #mixture of experts #open-weight models