Today's briefs
GTC Berlin 2026: Jensen Huang Keynote Covers Open Models, Sovereign AI, and Agent Tooling Across Europe
NVIDIA's GTC Berlin 2026 keynote from CEO Jensen Huang delivered a wide-ranging set of announcements centered on open models, sovereign AI infrastructure, and enterprise agent capabilities targeting European markets. Key themes included sovereign intelligence—enabling European organizations to run NVIDIA-powered AI on local infrastructure—alongside expanded cloud partner availability and new agent tooling. The breadth of announcements suggests NVIDIA is positioning itself as the full-stack AI platform provider, from chips to frameworks to deployment pipelines. For developers building agentic systems or deploying at scale in regulated industries, these announcements collectively affect the tooling and infrastructure layer. The keynote serves as the anchor event for the many individual GTC Berlin announcements below.
NVIDIA
NVIDIA Agent Toolkit Expands to Ground Enterprise Agents in Company Data
NVIDIA has expanded its Agent Toolkit with new capabilities specifically designed to connect enterprise AI agents to proprietary company data sources, announced at GTC Berlin 2026. The toolkit now provides tighter integration between NVIDIA's NIM microservices and enterprise knowledge bases, enabling agents to retrieve and reason over internal documents, databases, and structured data at inference time. This directly addresses one of the core production challenges for agentic systems: making agents reliably aware of organization-specific context without extensive fine-tuning. Developers building RAG pipelines or agentic workflows for enterprise clients can use this toolkit to standardize retrieval and grounding across deployments. The expansion positions NVIDIA's software stack as a serious competitor to cloud-native agent frameworks from Microsoft and AWS.
NVIDIA
Google Cloud and NVIDIA Announce General Availability of Vera Rubin Infrastructure
Google Cloud has announced the general availability of NVIDIA's Vera Rubin GPU infrastructure, bringing NVIDIA's next-generation compute architecture to GCP customers. Vera Rubin represents a significant leap in memory bandwidth and compute density over previous NVIDIA GPU generations, making it particularly relevant for large model training and high-throughput inference workloads. GA availability on Google Cloud means developers and ML engineers no longer need to wait for early-access programs to access this hardware tier. For teams training frontier-scale models or running inference at high concurrency, this is a meaningful infrastructure unlock. It also deepens the Google Cloud–NVIDIA partnership at the hardware layer, complementing the OpenAI–Google Cloud software-layer integration announced separately today.
NVIDIA
NVIDIA Cosmos Releases New Tools to Train, Evaluate, and Deploy World Models
NVIDIA has released new open tools under the Cosmos platform at GTC Berlin, expanding the ecosystem for building, evaluating, and deploying physical AI world models. World models are used to simulate environments for robot learning and autonomous systems, and the new tooling lowers the barrier for researchers and engineers to iterate on these systems. The additions include evaluation frameworks to benchmark world model quality and deployment utilities to move trained models into production environments. For developers working in robotics, autonomous driving, or simulation-based AI, this represents a meaningful infrastructure update to an already-open platform. The continued openness of Cosmos distinguishes it from proprietary simulation stacks and gives the research community a shared foundation to build on.
NVIDIA
Recursive Releases Flint: Open-Weight 40B MoE Coding Model with 2.5B Active Parameters
Recursive has released Flint, an open-weight 40-billion-parameter Mixture-of-Experts coding model designed specifically for agentic coding workflows, with only 2.5 billion parameters active per forward pass. The MoE architecture means Flint achieves strong coding performance while keeping inference compute and memory requirements comparable to a much smaller dense model, making it practical to run locally. It is explicitly designed for use in coding agents—not just autocomplete—meaning it is optimized for multi-step task execution, code editing across files, and tool use. For developers who need a capable, locally-runnable coding model without sending code to an external API, Flint is a notable new option in a space dominated by proprietary models. The open-weight release also allows fine-tuning on proprietary codebases, which is a significant advantage for enterprise and security-conscious teams.
MarkTechPost
Vercel Open-Sources durable-harness for Crash-Proof Coding Agents
Vercel has open-sourced durable-harness, a framework designed to make coding agents resilient to crashes, interruptions, and process failures during long-running agentic tasks. The core problem it solves is that coding agents executing multi-step workflows—editing files, running tests, calling tools—are highly vulnerable to state loss if the underlying process crashes partway through. durable-harness provides checkpointing and resumption primitives so an agent can recover from failure without restarting its entire task from scratch. For developers building or operating coding agents in production, this is a practical reliability primitive that addresses a well-known pain point not covered by most agent frameworks. The open-source release means teams can audit, extend, and self-host the harness within their own infrastructure.
MarkTechPost
OpenAI Builds Novel Ad Stack for ChatGPT, Expands to 26 New Countries
OpenAI has detailed the technical architecture of ChatGPT's native advertising stack, which it has now expanded to 26 additional countries after initial rollout. The ad system is built directly into the ChatGPT interface rather than layered on top, with a focus on user trust and query relevance—meaning ads are matched to conversational context rather than keyword targeting alone. The technical post explains how OpenAI designed the system to avoid degrading the user experience, which is a meaningful constraint that differs from traditional web ad systems. For developers building on top of ChatGPT or considering OpenAI's platform for consumer-facing products, the ad layer is a new variable to understand in terms of how responses may be influenced. The 100,000+ advertiser milestone signals this is becoming a significant revenue and product surface area for OpenAI.
OpenAI Blog
Perplexity and NVIDIA Introduce JOLT: Joint Teacher-Student Reinforcement Learning Method
Researchers from Perplexity and NVIDIA have introduced JOLT (Joint On-Policy Learning and Teaching), a new reinforcement learning method that trains student models using a joint on-policy process with a teacher model rather than purely offline distillation. The key innovation is that the teacher and student interact on-policy during training, allowing the student to learn from corrections in contexts it actually generates—rather than only from static teacher demonstrations. This addresses a core weakness of standard knowledge distillation, where student models often fail on distributions that differ from the teacher's demonstrations. For ML engineers working on efficient model training or distillation pipelines, JOLT offers a potentially more data-efficient and capability-preserving path to smaller models. The collaboration between a frontier search AI company and NVIDIA's research team signals growing investment in post-training efficiency research.
Perplexity
IBM Granite Speech 5.0 Adds Improved Transcription, Translation, and Speaker Attribution
IBM has released Granite Speech 5.0, an updated version of its open speech model with notable improvements in transcription accuracy, multilingual translation, and speaker attribution (diarization). The model is designed to run faster than its predecessor while improving word error rates across a range of languages and acoustic conditions. Speaker attribution is a particularly valuable addition for developers building meeting transcription, call analytics, or voice-based agent systems where knowing who said what is essential. Granite Speech 5.0 is available on Hugging Face, making it straightforward to integrate into existing pipelines. For teams evaluating open alternatives to proprietary speech APIs like Whisper or Google Speech-to-Text, this is a strong new benchmark to test against.
Hugging Face
OpenAI Agents Coordinate Global Vulnerability Disclosure in 'Patch the Planet' Initiative
OpenAI has described a coordinated vulnerability disclosure effort called 'Patch the Planet,' in which AI agents were used to uncover critical security vulnerabilities across global infrastructure and coordinate responsible disclosure with affected parties. The system demonstrates agentic AI operating in a high-stakes security context, autonomously identifying vulnerabilities and managing the complex multi-party communication required for responsible disclosure. For security engineers and developers building agentic systems, this is a significant proof-of-concept for AI agents handling end-to-end workflows in adversarial environments. It also raises important questions about the safeguards and oversight mechanisms required when agents take autonomous action in critical infrastructure contexts. The disclosure process reportedly involved multiple critical-severity findings, underscoring the practical capability of current-generation agents in offensive and defensive security tasks.
OpenAI Blog
How Specializing SLM Vocabulary Can Cut Inference Cost by 77% in Domain-Specific Applications
A new Hugging Face blog post from Dharma AI details VocabX, a technique for specializing the vocabulary of small language models (SLMs) to a target domain, reducing inference cost by up to 77% in real-world applications without retraining the model from scratch. The approach works by pruning the model's vocabulary to tokens that are actually relevant to the deployment domain, which reduces embedding table size and speeds up the softmax computation at inference time. For developers deploying SLMs in domain-specific contexts—legal, medical, coding, or other verticals—this is a practically actionable optimization that can dramatically reduce serving costs. The technique is particularly relevant for on-device or edge deployments where memory and compute are constrained. The post includes code and benchmarks, making it directly usable by engineers evaluating inference optimization strategies.
Hugging Face