Kimi K3: Moonshot's 2.8 Trillion Parameter Statement

Kimi K3: Moonshot's 2.8 Trillion Parameter Statement

Moonshot AI dropped Kimi K3 on July 16 — a 2.8 trillion parameter MoE model the company claims goes toe-to-toe with Claude and GPT-class systems on coding, reasoning, and knowledge work [4]. It's being called the largest open AI model ever released, with full weights landing July 27 [5]. Built with Huawei's backing and priced at roughly half of comparable OpenAI offerings, it's less a product launch than a geopolitical signal [6].

The US-China AI gap that everyone assumed was measured in years is now being measured in months, and Kimi K3 is Exhibit A. Developers on X are framing this as a repeat of DeepSeek's breakout moment — except this time nobody's pretending to be surprised.

For builders, the practical upside is real: a frontier-class model, fully open, at a fraction of the cost. For US labs charging premium prices on closed APIs, that's a direct threat to the business model, not just the bragging rights.

Mira Murati's Thinking Machines Ships Inkling

Thinking Machines Lab — Mira Murati's post-OpenAI venture — released its first model, Inkling, around July 15: a 975B parameter (41B active) multimodal MoE trained on 45 trillion tokens spanning text, image, audio, and video, with a 1M token context window [7]. It's Apache 2.0 licensed, paired with the Tinker platform for fine-tuning, and while it's not topping every benchmark, it's genuinely strong on reasoning and coding [8].

What matters here isn't raw scores — it's the bet Murati is making explicitly: customizable beats one-size-fits-all. That's a direct shot at the "one giant model to rule them all" strategy OpenAI and Anthropic have pursued, and it's coming from someone who helped build that strategy in the first place [9].

Combined with LeCun's Meta and Moonshot's Kimi K3, we now have three separate, well-funded bets that open-weights customization wins the next phase of AI. That's not a coincidence — it's a market signal.

Finland's NestAI Trains Sovereign Military AI With Estonia

Away from the frontier-model arms race, Finnish startup NestAI announced it has begun training frontier models specifically for battlefield autonomy and command orchestration, piloting with Finnish and Estonian Defence Forces using real operational data plus synthetic simulations [10]. The company — €100M raised, 200 engineers, founded just last year — is partnered with AMD, LUMI AI Factory, Patria, and Bittium, explicitly building for European sovereign capability rather than relying on US or Chinese infrastructure [11][12].

This is the Nordic AI sovereignty conversation moving from slideware to deployed pilots. Nordic and Baltic states have talked for years about not wanting to run critical defense infrastructure on someone else's model weights or someone else's cloud. NestAI training models on real operational data with two national defense forces is that conversation becoming policy.

Worth watching: this is exactly the kind of high-stakes, judgment-heavy application where "who orchestrates the AI" matters more than "who wrote the code" — coordination of unmanned systems and force decisions isn't something you vibe-code.