Kimi K3, third in the world at a quarter of Fable’s cost, and America opens the weights

🔗 Learn more about me, my work and how to stay in touch: maeste.it: personal bio, projects and social links.

A week where I return to a thread I’ve been weaving for a while, the race of open models, but with two pieces of news that made me say the scene is really changing. The first is Kimi K3 from Moonshot, a 2.8-trillion-parameter model that costs very little and, on the intelligence indices, ranks above Opus 4.8 and just below GPT-5.6. The second is the first open weight from a new American frontier lab, by Thinking Machines Lab, with a manifesto that asks for an AI that extends human judgment instead of replacing it. In the deep dive I focus on my thesis: the gap between the Chinese labs and the American frontier labs is closing fast, driven by the ruthless competition among the Chinese labs themselves. For us Europeans, every step forward by the open side is more ammunition for our insurance policy. On the podcast you’ll find “European inference and zero logs” with Eugenio Petullà, and episode #63 “American open weight and Kimi K3: a change of scene”. In the links: Google slowing down while the Chinese accelerate, Gemma 4 on the Pixel 10, Hassabis imagining a standards body, and a survey on self-improvement that speaks straight to my old harness fixation. Happy reading.

Podcast:

• On Wednesday “European inference and zero logs” came out, the interview with Eugenio Petullà (Regolo.ai) on LLM inference in Europe, zero data retention and model routing. Listen.

• Eugenio explains why your prompts on OpenAI probably end up in the logs, and why “data in Europe” isn’t enough if the provider answers to a government that isn’t yours. YouTube.

• On Saturday “American open weight and Kimi K3: a change of scene” came out (#63 Risorse Artificiali): Kimi K3 third in the world at a quarter of Fable’s cost, the first open weight from Thinking Machines Lab, a new American frontier lab, plus Hassabis on AGI and Torvalds on the kernel. Listen.

Anyone who follows the podcast already knows: this week I recorded an episode called “American open weight and Kimi K3: a change of scene”, and the title says almost everything. The open AI scene is changing, and fast. Let me lay it out.

I’ll start with the fact that struck me most. Moonshot launched Kimi K3, a 2.8-trillion-parameter multimodal model, a one-million-token context window, optimized for long-context decoding and for agentic coding. The open weights will arrive on July 27. But the numbers matter less than the positioning: K3 costs very little and, on the intelligence indices, ranks above Opus 4.8 and just below GPT-5.6. On the podcast I summed it up like this, the third model in the world at a quarter of Fable’s cost. I’ve tried it only a little, honestly, and not enough for a complete judgment, but the first impressions are excellent, especially on coding.

And here comes the part that really interests me. What impresses me isn’t one model in particular, it’s the speed at which the Chinese models are improving. DeepSeek, MiniMax, GLM, Kimi: every week there’s a leap, and the feeling is that the curve is getting steeper. My hypothesis is that what makes the difference is internal competition. The Chinese labs fight a war on price and performance in a much more ruthless way than the big American labs do, and the American ones are few, big and well positioned. More competition, more everything accelerates. And the result is there for everyone to see: the gap from the frontier labs is shrinking, and shrinking fast, at a pace that seems to be accelerating right now.

It’s the flip side of the argument I was making a few weeks ago, when Amodei was saying that open source is dangerous and must be limited. The more you try to close things down, the more the open side accelerates, and the more the open side accelerates, the more the insurance policy I keep talking about gets stronger. For us Europeans, having models at this level, free and low-cost, isn’t a technical detail: it’s our form of independence.

And here comes the news that closes the circle. So far, open weights have been a predominantly Chinese story. Now an American voice is added, and it’s an authoritative one. Thinking Machines Lab published a manifesto titled “The future worth building is human”, and released the first open weight from a new American frontier lab. I agree with the manifesto’s thesis: AI must extend human will and judgment, not replace them, and models shouldn’t be trained in a few centers and then frozen, but shaped by those who use them. There’s an explicit reference to Polanyi and Hayek, to tacit, local, distributed knowledge, which a centralized AI can’t grasp. It’s a decentralized and humanistic vision, and as it happens it’s the same direction I’m working in with harnesses and with Lince: giving users the lever to shape intelligence, not handing them a sealed box.

A change of scene, indeed. The open frontier is no longer only Chinese, and it’s no longer just a fallback to use when closed models get taken away from us. It has become the main game, with two accelerators: the ruthless competition among the Chinese labs and, now, a high-level American weight. The more the open side runs, the more my European thesis holds: we build on top, with harnesses and surrounding software, on top of models that nobody can take away from us.

The news is about Google, but I read it the other way around. Gemini 3.5 Pro is months behind precisely on coding, the ground where the agentic game is played, and while Google struggles the competitors pass it by. The detail that strikes me is in the same article: Moonshot and its Kimi get mentioned as a model that now rivals OpenAI and Anthropic. It’s no coincidence that in the deep dive I was talking about a shrinking gap. The American frontier is no longer alone, and at times it even struggles to keep the pace.

Google keeps betting on open and on local, and this Gemma 4 designed to run natively on the Pixel 10’s TPU is the proof. Offline chat, image recognition and audio transcription right on the phone, with no connection: it’s exactly the kind of on-device inference I talk about often. And since it’s open weight, it’s more ammunition for the insurance policy I was talking about in the deep dive. Capable, free models that also run in your pocket: for us Europeans it’s independence you can touch.

Hassabis proposes an independent body to oversee frontier models, with pre-release review and standardized benchmarks. On AGI safety the idea makes sense, and I talked about it on the podcast too. But I read it with the doubt I raised there: a body like this risks becoming a gatekeeper that rewards the big players and suffocates the very open weights I was talking about in the deep dive. It’s curious that exactly while Thinking Machines Lab asks for decentralized AI shaped by the user, on the other side they ask for more center. Two visions of the frontier pulling in opposite directions.

This paper isn’t directly connected to today’s deep dive, but it speaks my language. The authors, with Schmidhuber in the front row, describe the agent as a foundation model coupled with a scaffold of prompts, memory, tools and control logic, and they formalize self-improvement as an operator that updates the weights or precisely that scaffold. The scaffold is my harnesses, called by another name. It’s yet another confirmation of my old fixation: the lever to improve lies more and more in the software surrounding the model, not only in the model. Especially when, as I was saying in the deep dive, the model is open and you control it.