In a 24-hour window in the middle of July, two labs opened the weights on two large models with the same two headline numbers — open license and a 1-million-token context. On July 15 it was Inkling, the first model from Mira Murati's Thinking Machines Lab. On July 16 it was Kimi K3, Moonshot AI's 2.8-trillion-parameter flagship. The tech press filed them under the same story: China and a US lab both shipped big open models this week.
They are not the same story. They are close to opposite ones, and if you're a team of one deciding where to spend the next month, telling them apart is the whole game.
The one-screen answer#
Kimi K3 is a scale bet. Inkling is an ownership bet. K3 is the largest model ever opened — 2.8T parameters — and its near-term value to you is a cheap hosted frontier API, not a download. Inkling is deliberately not the strongest model available; its value is that it's a 975B base you fine-tune into your own model and keep. So the real question was never "which open-weight model." It's: do you want to rent a bigger generalist, or own a specialized base? Everything below is that sentence, expanded.
Kimi K3: rent the giant, self-host later (or never)#
Moonshot's number is the one everyone repeated: 2.8 trillion parameters, 896 experts, 16 active per token, a 1M-token window, full open weights promised July 27. It is a genuine milestone — the first open-weight model in the three-trillion class. But "open" is doing less work than it looks like. At roughly 1.4TB of MXFP4 weights, K3 is a multi-GPU or rented-inference deployment, not a side project. And at launch the "open" weights weren't published yet, nor was the Modified MIT license file — a promise, not a document.
So for now the honest use of Kimi K3 is the hosted API at $3/$15 per million tokens ($0.30 cached input) — a competitive price to touch frontier-scale capability with zero infrastructure. The benchmark story is the part to hold loosely: Moonshot's coding claims lean on newer suites that aren't yet independently replayable, with no SWE-bench Verified or Pro figures at launch. We laid out the full founder playbook — prototype on the API, measure your own tasks, self-host only when volume bends the cost curve — in what a founder actually does with Kimi K3.
Inkling: own a base, not a chatbot#
Thinking Machines went the other way on purpose. Inkling is a 975B mixture-of-experts model that activates only ~41B parameters per token, trained on 45 trillion tokens of text, image, audio, and video, released under Apache-2.0 with weights on Hugging Face on day one. Artificial Analysis called it the leading US open-weights model at release — but the company says plainly it is not the strongest model available, open or closed. That's not modesty; it's the pitch. Inkling exists to be specialized through Tinker, their customization platform. The deliverable is a base you own that answers your thing better than a rented generalist does. We walked the fine-tune-and-own-versus-rent-and-prompt decision in full in Inkling: the base a founder fine-tunes instead of renting.
One lab opened the biggest model and told you to rent it. The other opened a smaller one and told you to make it yours. Same week, opposite advice.
So which one is yours?#
Strip the launch noise and match the model to the shape of your product:
- **You need peak general reasoning now, at modest volume. Neither open model is your answer — rent a closed frontier model (Opus 4.8, GPT-5.6) and move on. But if you want a cheaper big generalist to prototype against, Kimi K3's API** is the move.
- Your product is a narrow slice repeated a million times. This is most products. A specialized Inkling fine-tune — 41B active, weights you host — can match a generalist on your slice while flipping the economics and the ownership in your favor.
- Data can't leave your infrastructure, or per-token rent is eating you at scale. Both point at Inkling today (weights already downloadable) and at Kimi K3 after July 27, once its weights and real license are in hand.
The trap is treating this as a capability contest and picking the bigger number. A 2.8T generalist will out-reason a fine-tuned 41B-active base on a random hard prompt — but "random hard prompt" is not a business. Pick the bet, not the parameter count: rent the giant when you need breadth this week; own the base when your edge is depth you can keep.



