Thinky's Twinky Inkling
Thinking Machines has released their first model with a spectacular comms push. The benchmarks are waifish but the direction is promising and there's much to love. USA! USA!
Perhaps because of its velocity, the AI discourse online has a tendency to be a little swing-y. OpenAI is building the singularity; no, the future is Claude. Actually, the future is kind of Chinese, maybe?
Well, Mira Murati’s Thinking Machines chose the perfect moment to release its first model: an open-source, open-weight model called Inkling. Open source is in right now, a fact partly attributable to Thinky itself, which released a paper with Bridgewater in June showing that a fine-tuned open model could outperform frontier closed models at a tenth of the price. Inkling is a general purpose and multimodal model that comes ready to be fine-tuned using Thinky’s Tinker.
I have never seen a better-orchestrated product release, or a more masterful one. TechCrunch covered it. Wired wrote about it. The FT covered it in salmon pink. NVIDIA tweeted about it twice, from two different accounts. Everyone is talking about it. Even Databricks is talking about it. And just about everyone spoke in glowing terms. Not because Inkling seems to be a particularly capable model—it performs in the middle of the pack on many benchmarks—but because of what it represents.
Narrative fatigue has begun to set in for the big labs. Anthropic releases an ad with the title: “There’s hope in hard questions”, featuring a man’s voice, filtered to sound like a call from a landline, asking “who’s gonna hit the brakes, if we need to?” over a static image of Arlington National Cemetery. Sam Altman quote-tweets it with a snipe, and follows up with a jab about Fable’s safeguards. Normal people don’t care about this, but if you spend a lot of time online, like I do, maybe you’d like an alternative? A human future, perhaps? Well here it is, and Wired and TechCrunch and the FT are all rooting for it.
Much of the technical praise online is for the direction. Researchers I spoke with, and others writing online, said that Inkling looks reasonably competent across a range of benchmarks. Its multimodal capabilities—that is, its ability to reason over audio and vision—are particularly good for an open model, thematically consistent with the interaction model Thinking Machines shared earlier in the year. It’s possible that multimodality specifically will open up interesting space for fine-tuning.
There are also signs that Thinky’s models will continue to improve faster than the frontier. Buried in the paper was a benchmarking comparison that showed Inkling-small—a lighter-weight, faster, cheaper model that Thinky plan to release soon—matching or outperforming the larger model. The Bridgewater result was achieved using Qwen3, but Inkling suggests that, if, for some reason, all Chinese open source models disappeared tomorrow, fine-tuned American models could still out-compete closed frontier models on specific tasks.
Which brings us neatly to the Chinese AI ecosystem. The Bridgewater paper was one reason for the open source moment; GLM 5.2 is another. GLM 5.2 is roughly the same total parameter size—~750b parameters to Inkling’s ~1T—but, before any additional fine-tuning, it is quite a bit more capable at writing code. There are not yet benchmarks for post-trainability, and it’s possible that the utility of a model as a base for specialization is not the same as its performance on general benchmarks. But if you were going to post-train a custom model tomorrow for an internal workflow, there is a strong case for starting with GLM 5.2.
(Inkling is multimodal, whereas the GLM model is not, out of the box. That said, Harry Partridge from Baseten did run a cool postraining experiment to give GLM 5.2 vision capabilities, so it’s not a sure thing.)
And in the coming weeks, we will likely see even more capable models from Chinese AI labs. Moonshot is currently teasing the release of its next model, Kimi K3. They even produced a tasteful short-form video. With jazz playing. The FT is reporting that K3 will outperform Opus 4.8 on benchmarks. Some of our more excitable friends on twitter are saying that it might be on par with Fable 5, based on early impressions from Arena and elsewhere. By the time this email reaches your inbox, it will likely be live, and you can decide for yourself.
There is, in any case, always another model. This is Thinking Machines’ first general purpose, open model release, but it won’t be the last. As you can already tell from the coverage of this one, there is certainly demand for what Thinking Machines is selling—a humanist vision of AI as a tool for people to do cool things with. Use the model, maybe even give it some fine-tuning. From their personal accounts, Thinking Machines staff often use the word “play”.
