Mira Murati's Thinking Machines Lab shipped its first open-weight model, Inkling, in mid-July 2026. Two weeks later, on July 31, the company followed up with Inkling-Small: a model at roughly a quarter the parameter count that beats the original on several benchmarks. That's an unusual thing to admit about your own flagship, and it's worth looking at why it happened.
The two models, side by side
Inkling is a mixture-of-experts model with 975 billion total parameters, though only about 41 billion are active for any given task. It was trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all four instead of bolting on separate encoders for each modality.
Inkling-Small cuts that down to 276 billion total parameters with 12 billion active per token, licensed under Apache 2.0. Fewer active parameters usually means a weaker model. Here, Thinking Machines says Inkling-Small actually surpasses Inkling on several benchmarks, not just approaches it.
| Inkling | Inkling-Small | |
|---|---|---|
| Total parameters | 975B | 276B |
| Active parameters per task | 41B | 12B |
| License | Open-weight | Apache 2.0 |
| Released | Mid-July 2026 | July 31, 2026 |
Why a smaller model beating the bigger one is the actual story
Most labs release a big flagship, then a smaller distilled version that trades some quality for speed and cost. What's different here is Thinking Machines saying the smaller model wins on parts of the benchmark suite, not just that it's "good enough for the price." That's a real signal about how much headroom was left in the training recipe for Inkling, separate from raw parameter count.
It also fits Thinking Machines' stated bet: that one-size-fits-all frontier models aren't the only path forward, and that smaller, more efficient, fully open models can compete on specific fronts. Both Inkling models are open-weight, meaning anyone can download the weights and fine-tune or modify them directly instead of working through an API with someone else's guardrails.
Pricing and how to actually try it
At launch, Thinking Machines is running a limited-time 50% discount on Inkling-Small's API pricing. You can fine-tune Inkling-Small on Tinker, the company's training platform, or try it without any setup in the Inkling Playground.
It's a similar playbook to what [Moonshot AI did with Kimi K3](https://questloops.com/blog/how-to-try-kimi-k3-for-free-moonshot-ai-s-2-8t-open-source-model) and what [DeepSeek did with V4 Flash](https://questloops.com/blog/deepseek-v4-flash-0731-explained-pricing-benchmarks-and-why-it-s-beating-its-own-flagship): open weights, a free or discounted way to try the model immediately, and a bet that developer goodwill matters as much as benchmark scores.
For anyone deciding which one to actually use: Inkling-Small is the more practical starting point right now given the discount and lower compute footprint for fine-tuning, and you'd only reach for the full 975B Inkling if you specifically need whatever extra headroom the larger active-parameter count buys you, which the benchmarks so far don't clearly show.
FAQ
**Is Inkling-Small actually free?** It's open-weight under Apache 2.0, so you can download and run it yourself at your own compute cost, and Thinking Machines is also offering a limited-time 50% discount on hosted API pricing.
**Where can I try it without installing anything?** The Inkling Playground, linked from Thinking Machines' site, lets you test the model directly in the browser.
