Cogent AI staff launched Cogent VR-1, a reasoning mannequin post-trained particularly for cybersecurity fairly than choosing up cyber functionality as a facet impact of normal coding power. It ships with two companions: IntrusionBench, a benchmark that scores brokers on accomplished enterprise intrusions, and the Cogent AI Harness, a ruled runtime for safety brokers. The launch…
Onton, a San Francisco-based search and discovery firm, has launched Ontology 1, a neurosymbolic mannequin for advanced, conversational, multimodal product search. On a 90-query benchmark scored by three unbiased LLM judges, Ontology 1 reached a imply precision@10 of 0.630, towards 0.543 for Google Shopping and 0.469 for Amazon. It did this whereas indexing roughly 1%…
In this tutorial, we design a whole GeoAI workflow for extracting constructing footprints from high-resolution NAIP aerial imagery. We start by configuring the geospatial deep studying setting, downloading raster imagery and vector labels, and inspecting their spatial properties earlier than producing georeferenced picture chips and segmentation masks. We then prepare a U-Net mannequin with a…
Thinking Machines Lab has launched Inkling-Small, an open weights Mixture-of-Experts mannequin with 276B complete parameters and 12B lively. That is a couple of quarter the dimensions of Inkling, which carries 975B complete and 41B lively parameters. The mannequin was educated on NVIDIA GB300 NVL72 methods. It causes natively over textual content, photos and audio. The…
Agentic reinforcement studying analysis is fixed algorithm modification. New estimators, new pipeline phases, new rollout schemes. In mainstream frameworks every change threads via layers of coach, distributed backend, and rollout glue. That price lands on the researcher at each iteration. Molt, from NVIDIA’s NeMo staff, targets that price straight. Its a PyTorch-native agentic RL framework…
In this tutorial, we construct a sophisticated end-to-end time-series forecasting workflow with TimesFM 2.5. We start by configuring the runtime, putting in the required dependencies, detecting out there {hardware}, and producing a sensible multi-store retail dataset with pattern, seasonality, pricing, promotions, holidays, temperature results, and random variation. We then load and compile the TimesFM 2.5…
AMD launched Instella-MoE-16B-A3B, a completely open Mixture-of-Experts language mannequin skilled from scratch on Instinct MI300X and MI325X GPUs. The mannequin holds 16B complete parameters however prompts solely 2.8B per token. AMD is publishing weights from each coaching stage, together with information mixtures, coaching configs, and inference code. Two systems-level decisions carry the discharge: Gated Multi-head…
In this tutorial, we discover how NVIDIA Transformer Engine accelerates transformer workloads by combining fused GPU kernels, BF16 computation, and hardware-aware FP8 execution. We start by putting in Transformer Engine and detecting the lively GPU structure in order that we are able to decide whether or not the runtime helps TE kernels, FP8 tensor cores,…
Supabase has open sourced Supabase Evals, its benchmark and framework for testing how nicely AI brokers construct utilizing Supabase. It runs coding brokers together with Claude Code, Codex, and OpenCode in opposition to actual duties, comparable to constructing a schema, debugging a failed Edge Function, or fixing a damaged RLS coverage, then scores the end…
MiniMax releases MiniMax H3, a general-purpose multimodal technology mannequin. MiniMax H3 shouldn’t be a text-to-video mannequin with add-ons. MiniMax describes it as a general-purpose multimodal technology mannequin that reads textual content, pictures, video, and audio as one unified context and returns video with native stereo sound. The mains specs embody: 2K output, 4–15 seconds, integer…