Kimi K3, Inkling, and the new race for open models: what actually matters for a company

In our last round-up, in mid-July, we wrote that the open-weight podium was changing almost every month. That held for a few days: between July 15 and 16, 2026, two releases landed that shifted the balance again — one from China, one from a newly founded American lab. For anyone deciding on an AI architecture in a company, the interesting question isn't which of the two is "the best," but what their licenses say and what that actually means for staying compliant.

Kimi K3: the largest open model ever released

Moonshot AI unveiled Kimi K3 on July 16: a Mixture-of-Experts model with roughly 2.8 trillion total parameters (just under 900 experts, with only 16 active per token), a context window up to 1 million tokens, and native vision capabilities. According to independent benchmarks cited by multiple outlets, it currently ranks fourth among all models — closed or open — trailing only the two most recent frontier models and ahead of established closed competitors. The market reaction didn't go unnoticed: shares of major semiconductor makers fell sharply in the following days, and more than one observer described it as a second "DeepSeek moment."

The license, though, isn't the "clean" MIT or Apache 2.0 you might expect from the headlines: Moonshot has long used a Modified MIT License, which freely allows commercial use, fine-tuning, and redistribution, but requires visibly displaying the model's name in the interface if the commercial product embedding it crosses 100 million monthly active users or $20 million in monthly revenue. For the vast majority of companies that threshold stays theoretical, but it should still be read before adopting the model, not assumed. Full weights are due by July 27, 2026: as we write this, Kimi K3 is announced and benchmark-verified, but not yet downloadable.

Inkling: Thinking Machines Lab's first open model

The day before, on July 15, Thinking Machines Lab — the lab founded by Mira Murati, OpenAI's former CTO — released Inkling, its first public model, and it's open-weight: roughly 975 billion total parameters in an MoE architecture, with only 6 of 256 experts active per token (about 41 billion active parameters), natively multimodal across text, images and audio, trained on 45 trillion tokens. The license is Apache 2.0, with no hidden thresholds or clauses: commercial use, modification, and redistribution are free from the first user onward.

It's a signal worth noting beyond the technical specs: a high-profile American lab, with substantial capital and no need to chase immediate revenue, chose openness as its first public product instead of a closed, API-only model — until now the norm among US labs, with Meta as the partial exception. Among the team's stated goals are keeping inference costs down and making the model more resistant to censorship.

Meanwhile, Alibaba isn't standing still

In the same week, Alibaba previewed Qwen3.8-Max, a 2.4-trillion-parameter model, announcing that this new generation's weights will also be released — continuing the strategy already seen with the 3.6 series, again under an Apache 2.0 license. It's yet another confirmation that, at this stage, the pace of open-weight releases is measured in days, not months.

What actually matters for a company deciding today

• Chasing this week's announcement isn't a strategy: at this pace, whatever model you pick today as "the best" risks not being that in a month. What matters is an architecture that lets you swap it out without rebuilding everything.
• Licenses remain very different from one another even when they all look "permissive": Kimi K3's Modified MIT has attribution thresholds that Inkling's and Qwen's Apache 2.0 doesn't. Read them one by one — don't assume from the label.
• The market's reaction to Chinese releases is a practical reminder, not just a financial one: an architecture built to be model-agnostic — where the model is swappable and your data stays under your control, on your own infrastructure or a Swiss sovereign cloud — protects you from the volatility of a sector that changes leaders every few weeks.
• AI Act deadlines don't wait for the next release: penalties for general-purpose models start applying from August 2026, days from now. Checking your architecture's compliance now costs far less than discovering a problem later.

This is the work we do every week with our clients: tracking these releases, reading the licenses before they become a problem, and building an architecture where the right model runs where it needs to run — with your data staying yours, regardless of who's leading the podium that week.