Nvidia gives away a capable agent model — weights, training data and all
On 11 August Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model that activates only 3 billion parameters per token. The weights are on Hugging Face and Nvidia's build platform, free for commercial use with no permission request and no licence fee, and Nvidia published the training data and recipe alongside them, subject to some licensing constraints. It is Nvidia's first open-weight model since CEO Jensen Huang publicly swung behind the open approach last month. The model is built for autonomous agents; Nvidia says it runs on a single consumer RTX GPU or a DGX Spark and claims up to 4x faster output and roughly 30% faster agentic task completion than comparable open models — vendor figures, not independent ones.
For a regulated European team, a downloadable agent model with published provenance is close to the profile they have been asking for: nothing leaves the building, and the training data is inspectable rather than a black box. Nvidia also shipped NeMo Switchyard, an open library that routes each agent request to the most efficient model for the task — an acknowledgement that most production stacks already run more than one model. Check the licence terms and re-run the speed claims yourself before committing.
OpenAI hands vetted defenders an offensive-security model
On 10 August OpenAI expanded its Daybreak security program and launched GPT-5.6-Cyber, a version of GPT-5.6 Sol tuned for offensive security work such as finding zero-days and building exploit chains. It is available only to vetted defenders through a new “Daybreak Red” tier, while a “Daybreak Blue” tier strips standard guardrails for defensive tasks like malware analysis and incident response. In OpenAI's own internal evaluation of exploit-chain, authentication-bypass and privilege-escalation requests, the cyber model completed 95% versus 1.5% for standard Sol, and OpenAI says it used the model to find two previously unknown bugs in V8, the JavaScript engine behind Chrome. All individual Daybreak accounts must adopt hardware security keys from 1 September. The bet is that arming defenders with frontier offensive capability under tight access controls beats leaving it locked away while attackers catch up — a call every enterprise security team now has to weigh for itself.
Mistral turns its sovereignty pitch into a product and a bank deal
France's Mistral used 11 August to make “sovereign AI” concrete rather than rhetorical: regionally locked inference with a priority tier for mission-critical workloads, wider access to third-party open models running on its own platform, and a coalition to help fund European compute capacity through 2030. Days earlier it signed Dutch bank ABN AMRO — its first tie-up with a major Netherlands lender — to build AI tools built and overseen inside Europe, part of the bank's push to lean less on providers based outside the continent. The useful signal for European buyers is the demand, not the supplier: regulated institutions are starting to treat in-region inference and provider provenance as procurement line items, not preferences.
Quick Hits
- June AI raises 0M to fix enterprise deployment. The New York startup came out of stealth with a 0M pre-seed led by Marc Benioff's TIME Ventures, with Michael Dell and Aaron Levie participating, to use agents to wire modern AI into legacy systems — a bet that the bottleneck is deployment, not the model.
- HappyRobot joins the unicorn club on supply-chain agents. The San Francisco-based, Spanish-founded firm raised a 50M Series C at a .2B valuation, led by Prysm Capital and co-led by Eurazeo, for voice-and-email agents that run logistics operations for customers including DHL and Kuehne+Nagel.
- Qwen's open-weight week arrives empty. Alibaba said Qwen3.8-Max and a new 27B checkpoint would hit Hugging Face “the week of 10 August”; as of today no repository has appeared and the licence is still undisclosed — verify the artifact, not the announcement.
- Agent pilots still stall before production. Fresh surveys keep hitting the same wall: Forrester and Anaconda research has roughly 88% of agent pilots failing to reach production, with evaluation and governance gaps the top blockers — though a Caylent survey (6 Aug) found 59.5% of leaders already run at least some agents in production.