EVA Tech BriefIT/DevOps

2026-06-25 - 7:30 PM PKT

Frontier LLMs

Databricks argues for open frontier AI stacks

Databricks leaders Matei Zaharia and Reynold Xin made the case that the next wave of enterprise AI will depend on open model, data, and agent-cloud ecosystems rather than closed end-to-end platforms. The discussion is useful for teams planning long-lived AI infrastructure because it frames openness as a way to preserve portability, observability, and control as agent workloads mature.

NVIDIA NeMo AutoModel targets faster transformer fine-tuning

Hugging Face highlighted NVIDIA NeMo AutoModel as a route to accelerate transformer fine-tuning workflows. The post is relevant for AI engineering teams looking to reduce experimentation time while keeping training pipelines closer to common open-source model tooling.

GPUs and Silicon

Broadcom and OpenAI detail custom Jalapeno inference ASIC

Broadcom and OpenAI unveiled a custom inference processor described as OpenAI's first dedicated chip effort, built on an unusually compressed nine-month cycle. The announcement points to continued vertical integration in AI infrastructure as major model providers try to optimize performance per watt and reduce dependence on general-purpose accelerators.

Qualcomm unveils HBC near-memory AI accelerator architecture

Qualcomm introduced its HBC near-memory architecture alongside AI250 and AI350 accelerators, positioning the design as a way around bandwidth and capacity limits in AI workloads. The company claims large gains in bandwidth per watt versus HBM and much higher capacity than on-chip SRAM, which would matter for dense inference deployments.

Cloud and Infra

Kubernetes WG Device Management focuses on AI and edge hardware

The Kubernetes project spotlighted Working Group Device Management, which is tackling hardware allocation for AI, edge, and telecom workloads. The work goes beyond CPU and memory scheduling to cover GPUs, TPUs, network interfaces, post-start allocation, and time-sharing, all increasingly important for production clusters running specialized hardware.

BreachRx launches AI incident command platform

BreachRx launched the Rex Platform, an agentic AI incident command center designed to coordinate response when multiple breaches or AI-accelerated attacks unfold at once. The product expands the company's earlier generative AI assistant into a broader workflow platform for incident coordination, evidence handling, and response orchestration.

Pakistan Tech

Security

Cisco SD-WAN zero-day gave attackers root access

Mandiant shared new details on exploitation of Cisco Catalyst SD-WAN vulnerability CVE-2026-20245, including how attackers created rogue root accounts on targeted devices. The reporting reinforces that edge and network management appliances remain high-value targets and need rapid patching plus account and configuration audits after disclosure.

Amadey and StealC infrastructure disrupted

A coordinated law enforcement operation with Bitdefender, Bitsight, ESET, Microsoft, and others disrupted infrastructure used by the Amadey and StealC malware ecosystems. Europol said the operation recovered 27 million stolen credentials and targeted criminal assembly lines used for ransomware, financial fraud, and critical infrastructure attacks.

Dev Tools

Mistral releases OCR 4 for structured document extraction

Mistral AI released OCR 4, a document intelligence model that returns structured document representations instead of just raw extracted text. It includes bounding boxes, block-type classification, and per-word confidence scores, making it more useful for enterprise pipelines that need auditable extraction from complex documents.

Passkey adoption gaps get public scorecard

A new website is naming major companies that still do not offer passkey support to users. TechCrunch reports that roughly a quarter of the world's most popular websites still lack passkeys, keeping password-based login risk on the table despite wider platform support for phishing-resistant authentication.