NVIDIA AI Updates: October 2, 2026

1. DOCA Agent Skills Give Coding Agents Verified BlueField APIs

NVIDIA. NVIDIA published DOCA Agent Skills, SKILL.md packages in the NVIDIA/skills GitHub repo that give general-purpose coding agents verified API signatures, hardware capability manifests, build-system specifications, and known failure modes for BlueField DPU development across the Flow, GPUNetIO, PCC, and RDMA libraries. In a 65-prompt evaluation, agents without the skills satisfied 19 percent of checklist items, while agents with them satisfied 100 percent across all prompts. A Go-based RDMA example built with the skills needed 73 percent less handwritten code and 46 percent fewer hardware commands. Source

2. DIN Deploy Ships C++ Samples for Local Speech, Segmentation, and Image Models on TensorRT for RTX

NVIDIA. NVIDIA released Do Inference Now (DIN) Deploy, an open GitHub collection of C++ samples that run local models through ONNX Runtime with the TensorRT for RTX execution provider on Windows and Linux, on both x86-64 and Arm64. The samples cover OpenAI Whisper, NVIDIA Parakeet TDT, and Nemotron ASR Streaming for speech, Meta SAM 2.1 for interactive image and video segmentation, and FLUX.2-klein-4B for image generation, with Vulkan and DirectX interop via ONNX Runtime 1.25. On DGX Spark, NVIDIA reports Parakeet TDT at 206x real time on GPU versus 14x on CPU, Whisper Large V3 Turbo at 58.5x versus 3.8x, and SAM 2.1 at 38.3 FPS versus 0.5 FPS. Source

3. Fine-Tuning Nemotron 3.5 ASR Cut Saudi Arabic Dialect Error Rate From 55 to 30 Percent

NVIDIA. NVIDIA documented a NeMo recipe for adapting Nemotron 3.5 ASR, a cache-aware FastConformer-RNNT streaming model, to Najdi and Hijazi Arabic using 133.7 hours of speech, with a replay mix of 90 percent Saudi audio, 7 percent FLEURS English, and 3 percent FLEURS Arabic to limit catastrophic forgetting. Word error rate on the Saudi test split fell from 55.05 to 29.96 percent and character error rate from 31.63 to 12.18 percent, while English WER moved from 11.04 to 10.42 percent. The fine-tuning notebook and related NeMo skills are on GitHub, and NVIDIA presents the minimal-curation recipe as a template for other low-resource dialects. Source