Thinking Machines' Inkling: an open-weight model you own
Mira Murati's lab shipped Inkling, the largest US-built open-weight model, under Apache 2.0. Here's why a downloadable brain matters for small operators.
Mira Murati's Thinking Machines Lab spent a year and a reported multibillion-dollar raise staying quiet. On July 15 it shipped, and the surprise isn't that the model is state-of-the-art — the lab says plainly it isn't. The surprise is the license. Inkling is an open-weight model under Apache 2.0, and per TechCrunch it's the largest US-built open-weight model you can download today. That's the part worth your attention.
What actually happened
Inkling is a mixture-of-experts model: 975 billion total parameters, but only about 41 billion fire on any given request, which keeps inference cost sane. It was trained on 45 trillion tokens across text, image, audio, and video, and handles a 1M-token context window. Full weights are live on Hugging Face, and it's wired for fine-tuning on Tinker, the lab's customization platform.
Thinking Machines is refreshingly blunt: Inkling is "not the strongest overall model available today, open or closed." The pitch isn't raw benchmark wins — it's efficiency and ownership. TechCrunch reports it hits equivalent coding performance using roughly a third of the tokens of Nvidia's Nemotron 3 Ultra, and a Bridgewater project scored 84.7% on financial reasoning at about one-fourteenth the cost of proprietary models. It's a starting point you fine-tune on your data, not a finished product you rent.
Why it matters for your business
A downloadable, Apache-2.0 model is a hedge against everything we keep warning about: price hikes, models getting pulled, and vendors you build on going to war. When the weights sit on your infrastructure, no one reprices your core capability overnight or deprecates it out from under you.
You don't need 975B parameters to act on this. The move is architectural: put a routing layer between your app and any model, so a frontier API and a self-hosted open model are both just config. Use the expensive closed model where it earns its keep, and a fine-tuned open one — trained on your corrections — everywhere else. Inkling makes that second path credible for more workloads than it was a week ago.
Key takeaways
- Thinking Machines shipped Inkling on July 15 — a 975B-parameter MoE (41B active), Apache 2.0, weights on Hugging Face
- The lab admits it's not the strongest model; the value is efficiency and ownership, not benchmark wins
- Claimed ~1/3 the tokens of Nemotron 3 Ultra for equivalent coding, and strong financial reasoning at ~1/14 the cost
- For operators, an Apache-2.0 model is a portability hedge — self-host it and no vendor reprices your core capability
Want AI that can't be repriced or pulled? We build model-agnostic systems with a routing layer, so a frontier API and a self-hosted open model like Inkling are a config change, not a rebuild. See how we build it or talk through your stack.
Sources: Thinking Machines — Introducing Inkling, TechCrunch — Thinking Machines' first open model.
- #open-weights
- #inkling
- #thinking-machines
- #model-portability
- #fine-tuning
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Pilot Protocol's $4.5M: a directory isn't a standard
A seed-stage AI agent network wants to be where agents discover each other. Useful, but know the difference between an open spec and someone else's front door.
Read itNvidia backs Safe Superintelligence, discloses no terms
Nvidia's Safe Superintelligence partnership names no dollar figure and no term length. Reported at $5B. Here's how to read AI deals that omit the numbers.
Read it