Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #30186

NeMo RL

Related event timeline, sources and context from the news index.

EVENT TIMELINE

3

MODELS · 1 SOURCE · NVIDIA Developer

NVIDIA releases Nemotron 3.5 Super VL post-training recipe and Litmus‑Bench V2.0 for scientific LLMs

NVIDIA published a practical recipe to post-train its open multimodal model Nemotron 3.5 Super VL for scientific tasks using NeMo Gym and NeMo RL, and released the Nemotron RL Litmus‑Bench V2.0 dataset and benchmark for molecular-property evaluation. In a drug‑metabolism example using MetxBioDB records, supervised fine-tuning reportedly improved accuracy by 12% and subsequent RL post‑training improved it by a further 17% under a 16,384‑token output budget; an optional distillation step using a stronger teacher model such as Claude Sonnet 5.5 is shown as a way to warm‑start RL.

7.0

COMPANIES · 1 SOURCE · NVIDIA Developer

NVIDIA Vera Rubin NVL72 posts up to 2x faster MLPerf Training v6.1 results vs GB300 NVL72

In its MLPerf Training v6.1 debut, NVIDIA's Vera Rubin NVL72 delivered the fastest time-to-train on six of seven benchmarks and achieved up to ~2x per-GPU performance versus an optimized GB300 NVL72 baseline; on the largest DeepSeek‑V3 671B workload Vera Rubin ran 1,024 GPUs in 4.14 minutes versus 8.03 minutes for the GB300 NVL72. NVIDIA also recorded the fastest absolute time on DeepSeek‑V3 using a 10,240‑GPU GB300 cluster (1.38 minutes); Vera Rubin maintained >97% efficiency scaling from 256→1,024 GPUs, completed the full seven-benchmark suite, and posted the best time for the new RL post-training Qwen 3.5 397B GRPO (10.44 vs 13.78 minutes), submitting results with two inference backends (TensorRT‑LLM and vLLM).

8.0

CODING · 1 SOURCE · NVIDIA Developer

NVIDIA's NCCL M2N Reshard implements Reshard collective to speed RL weight refit up to 5.7x

The Reshard collective reformulates RL "refit" as a DTensor-to-DTensor transfer that moves only the theoretical minimum data between arbitrary source and destination layouts. NVIDIA's open-source NCCL M2N Reshard reference implementation (nvidia/nccl-extensions) reports 2.7×–5.7× refit speedups in NeMo RL benchmarks on DGX H100 and DGX GB200 (tested on models including Qwen 3 235B and DeepSeek v3 671B), and uses a hybrid mode that applies Reshard to dominant feed-forward projection weights while retaining gather-and-broadcast for smaller tensors.

7.0