On July 15, an American AI company called Thinking Machines released its first large model, named Inkling. 975 billion parameters, able to understand images and audio, with all code weights published openly. But in its official announcement, the company wrote a line that left many readers stunned: “Inkling is not the strongest model currently available, open-source or closed.”
Most companies launching a new product would plaster “world’s number one” on their forehead. This one went the opposite way.
Then came the twist: a few hours after the announcement went out, it reached the top of Hacker News — 559 upvotes and 135 comments. The highest-voted comment in the thread read: “Don’t forget — it’s American. This is the first genuinely competitive non-Chinese open-source model since Llama 3.”
That contrast is worth talking about.
Image: The cover art Thinking Machines released for Inkling. Source: thinkingmachines.ai
The Two-Year Narrative of “China Leads Open-Source Models”
To understand why an announcement saying “not the best” set the tech world abuzz, you have to look at what happened over the past two years.
Between 2023 and 2025, the global open-source large-model landscape produced an awkward reality for Silicon Valley: the best open-source models were almost all coming from Chinese companies.
After Meta’s Llama 3 launched in April 2024, the US simply didn’t produce an open-source model that could truly go toe-to-toe with Chinese offerings in both performance and influence. Meanwhile, China’s Moonshot (Kimi K2.5 / K2.7), Zhipu (GLM 5.2), DeepSeek (V4 Pro), and Alibaba (the Qwen series) rolled out open-source models one after another, reshuffling the leaderboards several times over.
By the second half of 2025, “the future of open-source AI is in China” had become a widely discussed topic across the industry. The US side wasn’t idle — Google released Gemma, NVIDIA released Nemotron — but the community’s reaction was always “decent, but not at the Kimi level.”
So when Thinking Machines showed up with Inkling in July 2026, the fact that the “It is American” comment on Hacker News earned top votes itself reveals a psychological truth: the American tech community had been waiting for this day.
Who Is Thinking Machines?
The company’s founder is Mira Murati. If you follow the AI industry, you may have heard the name — she was formerly CTO of OpenAI and was deeply involved in the development of the GPT series. She left OpenAI in 2024 and founded Thinking Machines.
From the start, the company’s positioning differed from “closed-source giants” like OpenAI and Anthropic. Rather than chasing an all-powerful deity, they bet on a thesis: what enterprises truly need is a foundational model they can roll up their sleeves and modify themselves.
Inkling is the first product born from that thesis.
The Strategy Behind “Not the Strongest”
Inkling uses a Mixture-of-Experts (MoE) architecture — 975 billion total parameters, but only 41 billion are activated per inference. To borrow an analogy: it’s like a large company with 9,750 employees, but for any single task, only 410 need to show up to the meeting. The design aims to hold onto capability while controlling cost and speed.
It can process the equivalent of around one million English words at once (a 1M token context window), and its training data spans 45 trillion pieces of text, images, audio, and video.
On performance, according to third-party evaluator Artificial Analysis, Inkling scored 41 on the “Intelligence Index,” surpassing the previous best US open-source model, Nemotron 3 Ultra (38), making it the highest-scoring American open-source model to date. Thinking Machines’ own benchmarks show it beating the Chinese model Kimi K2.7 on several axes. That said, benchmark comparisons come with plenty of caveats — test methodology, evaluation criteria, and model versions all move the needle. Some community users reported that in real-world use, Kimi K2.7 still feels perfectly handy day to day.
Image: Thinking Machines’ performance comparison of Inkling against other open-source and closed-source models, published on HuggingFace. Source: huggingface.co
But the performance numbers aren’t the most interesting part of the Inkling release. What’s genuinely interesting is that Thinking Machines chose to admit it’s “not the strongest” — and put it in the announcement.
Why would a company volunteer weakness? My read is: they’re “drawing the battle lines.”
If you claim to be the strongest, the yardstick becomes those few rows of numbers on the benchmark leaderboard — the comfort zone of OpenAI, Anthropic, and Google, who burn billions every year just to lead on those few rows. But if you say “I’m not the strongest, but I let you modify, customize, and polish me into your own,” the yardstick changes. It’s no longer “who’s smarter,” but “who’s more obedient.”
In other words, Inkling’s real rivals are the open-source, self-deployable, fine-tunable models like Kimi, Qwen, and DeepSeek. And in that race, it chose to enter with a more humble posture.
An American Open-Source Comeback?
The community reaction converged on one point: the geopolitical significance.
Hacker News user paxys put it sharply: “This is the first competitive non-Chinese open-source model since Llama 3.” Another user, segmondy, added: “If the benchmark data is reliable, Inkling genuinely earns a spot on the shortlist for everyday use.”
There were dissenters too. Some pointed out that Arcee’s Trinity Large was also an American-team open-source model, but its marketing was so poor that most people never heard of it. Others brought up Google’s Gemma 4, arguing it deserves a place in the conversation.
But from the angle of community heat, Inkling’s launch did something its competitors didn’t: it made “American open-source AI” a topic again.
Behind this lies a larger narrative shift. Over the past two years, China’s rapid advances in open-source AI — especially the back-to-back releases from DeepSeek and the Kimi series — made “open source = China’s strength” almost a consensus. And now, a company founded by a former OpenAI core figure, with a humble posture that openly admits “not the strongest,” has pulled the conversation back to the American side.
Of course, a single model release changes nothing. Whether Inkling will actually be widely adopted by developers, or whether it gets overtaken again by Chinese models in later iterations, is unknown. But on July 15, 2026, one thing was certain: an open-source model built by an American company had once again reached the top of Hacker News.
And the next top-voted comment may already be on its way.
References:
- Thinking Machines: Introducing Inkling
- HN discussion (item?id=48924912)
- Artificial Analysis: Inkling debuts at 41
- TechCrunch: Thinking Machines amps up its bet against one-size-fits-all AI
- Axios: Mira Murati’s Thinking Machines debuts first AI model