Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released Inkling on July 15—a 975-billion-parameter multimodal model it trained from scratch and shipped with full open weights under an Apache 2.0 license.

What are we talking about?

• Open weights — the trained model’s numbers (its “weights”) are published, so anyone can download it, run it on their own machines, and change it.
• Apache 2.0 — a permissive license that lets you use, modify, and redistribute the model commercially, with very few strings attached.
• Mixture-of-Experts (MoE) — the model is split into many small specialist sub-networks; each word only wakes up a few of them, so just 41B of its 975B parameters run at a time.
• Multimodal — Inkling reads text, images, and audio directly, rather than having those senses bolted on afterward.
• Tinker — Thinking Machines’ fine-tuning service, where developers can adapt Inkling to their own data.

How it works

1. Thinking Machines pretrained Inkling on 45 trillion tokens of text, images, audio, and video using NVIDIA GB300 NVL72 systems.
2. Its Mixture-of-Experts transformer holds 256 routed experts plus 2 shared ones, activating 6 per token—so roughly 41B of 975B parameters fire at once.
3. Post-training used large-scale reinforcement learning across more than 30 million rollouts, plus a “controllable thinking effort” dial so users can trade tokens for accuracy.
4. The company published the weights on Hugging Face (including an NVFP4 checkpoint for NVIDIA Blackwell chips) and on its Tinker API.
5. To show what customization looks like, the lab says it had Inkling write its own fine-tuning job, run it, and grade the result.

What changed

Why it matters

• A frontier-scale open base. Companies can run and fine-tune a large multimodal model on their own infrastructure instead of renting a closed API.
• Multimodal from the ground up. Inkling handles text, images, and audio natively, which Thinking Machines wants as the reasoning core of its real-time “interaction models.”
• Efficiency is a selling point. The company says Inkling can match Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens, and it is offering Tinker access at a 50% discount for a limited time.
• Broad deployment. Inkling is already served through TogetherAI, Fireworks, Modal, Databricks, and Baseten, with transformers and vLLM support.

What it does NOT prove / caveats

• Company-reported benchmarks. The scores come from Thinking Machines, run at “effort 0.99.” Its SWE-bench numbers use an internal bash-only harness while rivals are self-reported—not an independent head-to-head.
• Not the best model. Thinking Machines states plainly that “Inkling is not the strongest overall model available today, open or closed”; it pitches it as a customization base.
• Inkling-Small is only a preview. The lighter 276B model (12B active) is shown but its full weights are not released yet.
• Open weights cut both ways. The lab reports strong safety scores but acknowledges it is still studying how fine-tuning on Tinker can weaken safeguards.
• Some results use other checkpoints. Forecasting and “epistemics” figures were measured on a different checkpoint than the released model.

End-to-end recap