Mistral Bets Big on Open Weights With 'Le Chonk'

Mistral Bets Big on Open Weights With 'Le Chonk'

Mistral previewed Large 4 on October 6 — a 1-trillion-parameter, natively multimodal MoE model (49-52B active params) trained on 3,800+ Nvidia Grace Blackwell GPUs sitting in European data centers [4][5][6]. It handles 160+ languages, 1M context, and is tuned hard for coding, agentic workflows, cybersecurity, and finance. Full open weights land by end of October under a custom license; preview API pricing sits around $1.36/$4.18 per MTok [4][5].

This is Europe's clearest "we can compete at the frontier and keep it open" statement yet. Mistral isn't just releasing a strong model — it's releasing a strong model trained entirely on European infrastructure, which matters enormously for sovereignty-conscious enterprises and governments who don't want their inference routed through US clouds [5][6].

The reaction on X has leaned heavily into the sovereignty angle — less "is this better than GPT" and more "finally, a frontier-class model we can actually self-host and audit." For Nordic and EU buyers specifically, that's the pitch that matters.

Nous Research Raises $90M to Push Hermes Into the Enterprise

Nous Research closed a $90M Series B at a $1.5B valuation on October 7, led by Robot Ventures with Nvidia, Samsung, YC, and USV participating — total funding now ~$160M [7][8][9]. The money goes toward taking Hermes, their open-source AI agent, from hobbyist/developer darling (24M+ clones since February, ~2.5% of global token volume) into something businesses can actually deploy: SSO, permissions, auditing, the unsexy enterprise plumbing [8][9].

The core stays MIT-licensed — CEO Dillon Rolnick is explicitly positioning Hermes as the decentralized, user-aligned counterweight to closed platform agents [9]. Big tech backing an open-source agent company is itself notable; Nvidia and Samsung aren't writing checks here for ideology, they're hedging against platform lock-in from the labs they also depend on.

Worth watching: this is the second story today (alongside Mistral) that's really about who controls the agent layer, not who has the best model. That's the actual battleground now.

Finland Halts Google's €13B Data Center Buildout

Finland's Supervisory Agency ordered Google subsidiary Tuike Finland to stop preparatory site work — tree felling, topsoil removal, road building — at two planned data center sites in Muhos (~330 ha) and Kajaani (~200 ha), pending completed environmental impact assessments [10][11][12]. Police are investigating suspected illegal logging at the sites. This is part of Google's record €13B European infrastructure push, and Google has acknowledged the shortfall and says it will comply [11][12].

This lands right as the EU AI Act's transparency enforcement phase kicks in (August 2026), and it's a useful reality check for anyone assuming Nordic countries will rubber-stamp AI infrastructure because of the jobs and tax revenue. Environmental review isn't a formality here — it has teeth, and it can stop a hyperscaler mid-construction [10].

For Nordic AI builders, this is a reminder that the region's appeal — cheap renewable power, political stability, strong rule of law — cuts both ways. The same institutions that make the Nordics attractive for data centers will also slow you down if you skip steps.