NVIDIA AI Brings Nemotron-3-Nano-30B to NVFP4 with QAD

Here is your today’s AI Dev Brief from Marktechpost, covering core research, models, infrastructure tools, and applied updates for AI developers and researchers.

NVIDIA’s Nemotron-Nano-3-30B-A3B-NVFP4 is a 30B parameter hybrid Mamba2 Transformer MoE reasoning model that runs in 4 bit NVFP4 while keeping accuracy close to its BF16 parent. The model uses NVFP4 weights, FP8 KV cache and a small set of BF16 layers, giving about 2 to 3 times higher arithmetic throughput and about 1.8 times lower memory than FP8, with about 3.5B active parameters per token and support for context up to 1M tokens. Accuracy loss from plain NVFP4 post training quantization is reduced using Quantization Aware Distillation, which trains the NVFP4 student to match a frozen BF16 teacher via KL divergence instead of replaying the original SFT and RL pipeline. On math and coding benchmarks, QAD recovers performance to within a few points of BF16, making NVFP4 and QAD a practical approach for deploying RL heavy reasoning models on NVIDIA hardware..… Read the full analysis/article here.

LingBot World, released by Robbyant from Ant Group, is an action conditioned world model that turns text and control inputs into long horizon, interactive video simulations for embodied agents, driving and games. Built on a 28B parameter mixture of experts diffusion transformer initialized from Wan2.2, it learns dynamics from a unified data engine that combines web videos, game logs with actions and Unreal Engine trajectories, with hierarchical captions that separate static layout from motion. Actions enter the model through camera embeddings and adaptive keyboard adapters, which are fine tuned while the visual backbone stays frozen. A distilled variant, LingBot World Fast, uses block causal attention and diffusion forcing to reach about 16 frames per second at 480p on 1 GPU node with under 1 second latency, and achieves leading VBench scores with strong emergent memory and structural consistency....… Read the full analysis/article here.

▶ [Open Source] Rogue: An Open-Source AI Agent Evaluator worth trying Codes & Examples

▶ How to Build Memory-Driven AI Agents with Short-Term, Long-Term, and Episodic Memory Codes Tutorial

▶ A Coding and Experimental Analysis of Decentralized Federated Learning with Gossip Protocols and Differential Privacy Codes Tutorial

150+ more open codes/notebooks here ➡️