Two weeks before the European Commission's AI Act enforcement powers go live for general-purpose models, the open-weights frontier keeps moving east. Here's what mattered for European builders over the last two days.
Moonshot announces Kimi K3 — the largest open-weights model yet
Moonshot AI began rolling out Kimi K3 on 16 July: a mixture-of-experts model at roughly 2.8 trillion total parameters (16 of 896 experts active per token) with a 1-million-token context window, aimed squarely at long-horizon coding and agent workloads. Two variants ship first — K3 Max for chat and agents, K3 Swarm Max for parallel processing — via the Kimi app and Kimi Code, with full weights promised on 27 July under a modified MIT license.
The European angle: once the weights land on Hugging Face, frontier-class capability becomes hostable under EU jurisdiction — the sovereignty case that made Kimi K2 and DeepSeek attractive applies again, at a larger scale. The catch is that serving a 2.8T MoE means multi-node GPU clusters that only a handful of EU inference providers can stand up. Watch who picks it up — and read the modified-MIT terms before you build on it.
EU AI Act: GPAI enforcement powers activate on 2 August
The grace year is nearly over. The general-purpose AI obligations have applied to new models since August 2025, but on 2 August 2026 the Commission's enforcement powers activate: information requests, model evaluations, mandated mitigations and recalls, with penalties up to €15 million or 3% of global turnover. Models placed on the market before August 2025 have until 2027 to comply. Providers who signed the General-Purpose AI Code of Practice and can show good-faith implementation get the collaborative treatment; everyone else gets the audit.
If you deploy GPAI-based systems in the EU, the two-week checklist is short: confirm your model suppliers have their transparency documentation in place (training-data summaries, copyright policy, downstream documentation per the Code of Practice structure), and if a supplier hasn't signed the Code, ask them why not — their answer is now a procurement signal.
Thinking Machines ships Inkling, its first model
Mira Murati's Thinking Machines Lab released its first broadly available model on 15 July, called Inkling — a multimodal model positioned on cost-for-performance rather than leaderboard peaks. It's the first shipping evidence for one of the most-watched (and best-funded) new labs.
For European buyers the placement is better than the frontier norm: Inkling shipped with full open weights on Hugging Face at launch — 975B total parameters with 41B active, Apache 2.0, no field-of-use restrictions — alongside the hosted API. That takes it out of the pure CLOUD Act / FISA 702 bucket: 41B active parameters is tractable on a modest H100 cluster, so a European operator can run it entirely inside its own jurisdiction. Capability is not the constraint here, and for once neither is jurisdiction — the constraint is standing up the serving capacity.
Enforcement begins in sixteen days. The suppliers who did the paperwork will be easy to spot.
Correction (19 July): an earlier version of this brief said Inkling was available as a US-hosted API only and drew a jurisdictional conclusion from that. Inkling's weights were published on Hugging Face under Apache 2.0 at launch; the story above has been updated.