Nvidia is placing a major bet on open-weight models to challenge traditional frontier systems. Speaking during his keynote at Nvidia GTC Berlin 2026—just days after the rollout of the Nemotron 3.5 Super VL open model—Jensen Huang highlighted how targeted enterprise customization can elevate open models to elite performance tiers, pointing to an initiative where AI21 Labs helped security firm Alice turn Nemotron into a system that surpasses standard frontier capabilities.

The collaborative venture involved AI safety specialist Alice teaming up with AI21 Labs and Inflection AI to optimize the safety metrics of Nvidia’s Nemotron architecture. Through their partnership with AI21, Alice successfully post-trained Nemotron 3.5 Super VL, boosting its defense against indirect prompt injection by 41 points to rival GPT-6.1 Sol. Simultaneously, Alice assisted Inflection AI in building a red-teaming dataset designed to minimize sycophancy in a fine-tuned iteration of Nemotron 3 Super 120B-A12B, ultimately lifting the alignment score of the Pi conversational assistant by 30%.

Overall, this initiative highlights how strategic post-training on curated datasets enables open models to punch above their weight and rival frontier-level efficiency across targeted applications.

How Open Models Can Beat The Frontier

As open-weight frameworks evolve into sophisticated, cost-effective alternatives, enterprise uptake has surged. Vercel’s AI Gateway, which directs trillions of tokens between production environments and AI developers, revealed that open models captured the majority of gateway traffic for the first time in August 2026, currently standing at 71.5% of all routed tokens.

Ori Goshen, CEO and co-founder of AI21 Labs, emphasized in a video interview that open frameworks have evolved over the past three or four months into legitimate contenders against proprietary systems for most workloads. As enterprise token expenses spiral, organizations increasingly demand deeper oversight of their AI infrastructure.

When you rent the AI, you actually pay twice. You pay for the tokens, and you pay for the missed opportunity to improve your AI system.

Goshen explained that true model ownership allows enterprises to gather telemetry traces, isolate system flaws, and continuously refine the underlying intelligence over time. Headquartered in Tel Aviv and founded in 2017, AI21 Labs operates an AI-labs-as-a-service model, embedding specialized researchers directly into client organizations. Beyond its work with Alice, the startup has partnered with global payroll leader Deel and revenue intelligence firm Gong to post-train Nemotron 3.5 Super VL for domain-specific supremacy.

  • Deel Partnership: AI21 post-trained Nemotron 3.5 Super VL to handle payroll inquiries, driving accuracy to 72%—surpassing Claude Opus 5.5 by 8 points while running at 33 times lower cost per query.
  • Gong Integration: The customized model matched Claude Opus 5.5 in both speed and extraction accuracy at a fraction of traditional operational expenditures.

AI security remains a primary concern amid the accelerated development race, with prompt injection standing out as a critical vulnerability. In these attacks, malicious users embed hidden instructions to bypass safety guardrails. Mitigating this risk was the core focus for Alice in its collaboration with AI21 Labs.

Noam Schwartz, CEO and co-founder of Alice, noted that the initiative aimed to demonstrate how effortlessly enterprises leveraging open-weight structures can attain frontier-grade security without incurring the massive expenditures required by proprietary labs.

Alice and AI21 targeted multi-step, sophisticated indirect prompt injection attacks. Schwartz explained that neutralizing these threats relies heavily on data volume and quality, noting that the team successfully enhanced model resilience with minimal data footprints through Nvidia's guidance.

Achieving superior outcomes with open models hinges on narrowing the use case scope. Schwartz pointed out that specialized deployments do not require solving abstract mathematical dilemmas or discovering new pharmaceuticals; they simply need flawless execution within a designated operational domain.

Established in 2018, Alice has collaborated with Nvidia and eight of the top ten foundation model creators, deploying proprietary threat-detection intelligence while securing $280 million in funding to date.

Inflection AI Looks To Misuse

Inflection AI and Alice pursued a distinct objective: adapting Nemotron to mitigate sycophancy and curb unhealthy user attachments over extended dialogues. Recognizing that prolonged interactions can heighten emotional vulnerabilities, Inflection partnered with Alice to post-train Nemotron 3 Super.

During the project, Alice generated over 20 distinct personas representing both vulnerable and adversarial users, deploying them in rigorous red-teaming simulations to refine system alignment. The resulting fine-tuning lifted Pi's alignment score by 30% while holding unnecessary response refusals down to 1.6%.

Sean White, CEO of Inflection AI, stressed that safety extends far beyond national security and baseline alignment to encompass protections for vulnerable demographics and long-term user safety—an area previously underexplored in multi-session interactions.

Inflection AI, launched in 2012 by Reid Hoffman, Mustafa Suleyman, and Karén Simonyan, powers its personal intelligence assistant Pi using proprietary custom-tuned models. The firm has collaborated extensively with industry giants like Meta and Nvidia while serving as a leading contributor to PyTorch.

With precise fine-tuning, open-source architectures prove formidable competitors to closed platforms. Mahmoud Ramin, research director at Infotech, observed via email that open models do not need universal supremacy to deliver immense enterprise value.

An open model does not need to become the best model at everything. It can become exceptionally good at a really valuable task.

Ramin noted that these breakthroughs indicate safety can be actively trained into behavioral patterns rather than merely dictated through restrictive rules, making open models highly attractive for high-volume, data-sensitive corporate environments. Nevertheless, he anticipates proprietary models will retain a vital foothold for organizations lacking the internal scale and technical expertise to maintain custom stacks.

Ultimately, the true competitive edge of open models lies not just in ownership, but in synthesizing proprietary corporate data, expert research labs, and continuous operational feedback tailored precisely to an organization's ecosystem.