The Story
Mira Murati’s Thinking Machines Lab has been one of the quieter big names in AI. Two billion dollars in funding, a founder who used to be OpenAI’s CTO and now runs the lab as CEO, and for a long time not much you could actually download and run. That changed on July 15, 2026, when the lab shipped its first open-weights model, Inkling. And then, barely two weeks later, on July 30, it dropped a follow-up called Inkling-Small.
That efficiency argument has a hardware mirror: why memory, not compute, is now AI’s ceiling.
For the flip side — capability judged by hard problems, not leaderboards — see OpenAI’s Astra solving open math problems.
Here’s the thing that makes Inkling-Small worth paying attention to. It isn’t a bigger, flashier model. It’s a smaller one. And it’s nearly as good as the original.
Let’s put numbers on that. The first Inkling is a Mixture-of-Experts model — a design where the network is split into many “expert” subnetworks and only a handful fire for any given token, so you get a huge model on paper but pay for only a slice of it at run time. Inkling has 975 billion total parameters but activates only 41 billion per token. Inkling-Small keeps the same trick but shrinks the whole thing to 276 billion total parameters with just 12 billion active. That’s roughly a quarter of the size on both counts.
The intuition most people carry around is that smaller means dumber. Not here. On Humanity’s Last Exam — a brutal reasoning benchmark built to resist memorization — Inkling-Small scored 31.6%, actually edging out the full Inkling’s 29.7%. It also posted 80.2% on SWEBench Verified (real GitHub bug-fixing tasks, versus the full Inkling’s 77.6%) and 89.5% on GPQA Diamond (graduate-level science questions). The lab is upfront about the trade-off: the small model pulls ahead on reasoning and agentic coding, while the big one still knows more raw facts and hallucinates a little less. So it’s not free. It’s a swap — you give up some breadth of knowledge to get a model that reasons just as hard for a fraction of the compute.
Both models are multimodal in the same way. They take text, images, and audio as input, reason natively across all three rather than bolting on a separate vision module, and handle a context window of up to 1 million tokens. Both ship under an Apache 2.0 license, which is about as permissive as open licenses get — you can use them commercially, modify them, and build products on top without asking anyone. Full weights are on Hugging Face. Fine-tuning runs through the lab’s own Tinker platform, and there’s a chat playground if you just want to poke at it.
That Tinker connection matters more than it looks. The original Inkling wasn’t pitched as a model you just call and forget. The lab describes it as “a broad, balanced foundation model” — deliberately not the strongest at anything, but flexible enough to adapt. The idea is that you take it, fine-tune it on your own data through Tinker, and end up with something specialized that you own. Inkling was the foundation. Inkling-Small is the same foundation, cheaper to build on. When you frame it that way, the small model isn’t a downgrade — it’s the version most people were probably going to want anyway.
One more detail that’s easy to skim past: the effort dial. Both models let you turn reasoning up or down — spend more tokens thinking on a hard problem, or answer fast and cheap on an easy one. Pair that controllable effort with 12B active parameters and you get something well-suited to the unglamorous jobs that eat most of a real AI budget. Coding tasks. Using an LLM to grade other outputs. Generating synthetic data to train yet another model. These are the workloads where latency and cost matter far more than squeezing out the last percentage point on a leaderboard, and they’re exactly where Inkling-Small is aimed.
Then there’s the part that usually gets buried: price. Running Inkling-Small costs $1.20 per million output tokens, versus $4.05 for Inkling. That’s not a rounding error. For anything you run at volume — grading outputs, generating synthetic training data, batch coding tasks — you’re looking at more than a 3x cost difference for a model that reasons about as well. The lab trained it on NVIDIA’s GB300 NVL72 systems using, in its own words, substantially less compute than the original.
The Takeaway
I keep coming back to a pattern this fits into. I wrote about OpenAI cutting GPT-5.6 Luna’s price by 80%, and earlier about Stanford’s work on predicting an LLM’s training bill before you spend it. Different companies, different angles, same underlying story: the frontier of AI right now isn’t “how big can we go.” It’s “how little can we spend to stay this good.” Inkling-Small is that idea turned into a product you can download.
What’s genuinely interesting is the strategy underneath it. Thinking Machines is deliberately not chasing the single strongest model on the leaderboard. Even the original Inkling, the lab admits, isn’t the best model available. What it’s optimized for is being a great starting point you customize — efficient, multimodal, and available for fine-tuning on Tinker. Murati has framed this as a bet against “one-size-fits-all AI.” The pitch isn’t “our model beats GPT.” It’s “download our model, fine-tune it on your data, and you own the result.” Inkling-Small sharpens that pitch, because a cheaper base model is a cheaper foundation to build your own thing on.
That connects to something I find quietly important about the open-weight direction generally. When the weights are yours under Apache 2.0, the model stops being a service you rent and becomes an asset you keep. You can run it on your own hardware, tune it in private, and it doesn’t disappear when a company changes its API terms or sunsets a model you built a product around. For a lot of teams — especially ones handling data they can’t send to someone else’s cloud, like health records or internal code — that’s the whole game. And a 276B model that activates only 12B parameters is small enough to actually deploy without renting a data center, which is the difference between “open in theory” and “open in practice.”
It’s worth naming who this is competing with. On the closed side, you’ve got the OpenAI and Google APIs — enormously capable, but rented, and priced by the token forever. On the open side, the obvious rivals are the big open-weight families from Meta and the Chinese labs like DeepSeek and Qwen, which have been setting the pace on efficient MoE models for a while now. Thinking Machines walking in with a 12B-active model that trades punches with its own 41B-active flagship, under Apache 2.0, at a third of the price, is the lab planting a flag in that second camp. And doing it two weeks after its first-ever release is a way of saying it intends to move fast there.
So what does this mean if you’re not training models for a living? Mostly it means the floor is rising. The cost of “good enough to build on” keeps dropping, and the good-enough models keep getting easier to own outright. The two-week gap between Inkling and Inkling-Small also says something about pace — this lab is iterating fast, and the efficient variant, not the flagship, was the quick follow-up. My read is that’s a signal. The interesting competition in AI is shifting from who has the biggest model to who can hand you the most capable one you’re allowed to walk away with.
Worth watching whether the rest of the field reads it the same way.
This article is for informational purposes only.
Photo: Growtika / Unsplash
댓글 남기기