DeAI Daily Brief — 10 September 2026

Today in DeAI: the Institute of Foundation Models shipped the K2 Horizon open-weight family up to 375B parameters, Mistral published the sovereignty manifesto behind its €3 billion raise, and the World Economic Forum framed open-weight AI as owned infrastructure.

IFM releases K2 Horizon, open weights from 0.9B to 375B

The Institute of Foundation Models released K2 Horizon, six Apache 2.0 models from a 0.9B dense model to a 375B-parameter mixture-of-experts with a native 512K context window. Weights are on Hugging Face now; IFM promises training data, recipes, code, and intermediate checkpoints to follow. Why it matters: if the data and code actually ship, this becomes one of the few frontier-adjacent families outside teams can reproduce rather than just download — every benchmark figure is currently IFM's own and unverified. (Hugging Face) — our full coverage

Mistral's manifesto behind the €3B raise

Mistral published "sovereign, open-weight AI as the technology frontier," the strategy statement accompanying its €3 billion Series D at a valuation above €21 billion, led by Samsung Electronics with Scaleup Europe Fund (EQT) and PSG Equity. The pitch is that enterprises and governments want frontier performance without surrendering control of infrastructure and the intelligence loop. Why it matters: this is the explicit thesis the round is priced on — shipped weights and signed sovereign contracts will test it, not the funding itself, and the sovereignty framing remains Mistral's own claim. (Mistral AI) — background: open-weight vs open-source models

WEF: open-weight AI as owned infrastructure

The World Economic Forum published a framing of open-weight AI as turning "rented access into owned infrastructure," arguing that downloadable weights shift AI from a subscription dependency into an asset organizations control. Why it matters: when a body like the WEF adopts the owned-infrastructure framing, the sovereign-AI argument Mistral is selling moves from vendor pitch to policy vocabulary — relevant to anyone tracking procurement and jurisdiction requirements. (World Economic Forum)

Confidential computing hits the CISO agenda

SC Media's 2026 briefing on confidential computing puts hardware-attested trusted execution environments in front of security leadership as a mainstream procurement item rather than a research topic. Why it matters: TEEs are the mechanism behind the "verifiably private" inference claims this publication treats as policy statements until attestation is validated — CISO-level attention means those claims start getting checked. (SC Media) — background: confidential AI inference and TEEs

Watching tomorrow

Whether IFM's K2 Horizon pretrain and midtrain dataset repos go public — that converts a strong weights release into a reproducible one and is the first real test of the "fully open" claim.