Nemotron 3.5 Lightning Free

Shattered's analysis confirms Nemotron 3.5 Lightning launched on August 11, 2026 as a 30B MoE model with 3B active parameters per token, designed to run on a single GPU. NVIDIA claimed up to 4x the output speed of similarly sized open models, and the model pairs with a companion routing tool called NeMo Switchyard for agent workloads. The article details a hybrid Mamba-2, attention, and MoE architecture, a 1-million-token context window, and five separate checkpoints including a BF16 reference build and an NVFP4-quantized version. Weights, training recipes, and data details ship publicly under the permissive OpenMDW-1.1 license for free commercial modification and redistribution.

Tech Insider reports NVIDIA shipped Nemotron 3.5 Lightning on August 11, 2026 as an open-weight 30B-A3B Mixture-of-Experts model with only 3B parameters active per token, formally positioned as the newest entry in the Nemotron lineup. The model targets always-on, narrow agent tasks with a 1-million-token context window, enough to hold a full codebase in memory. NVIDIA published checkpoints in BF16, FP8, and NVFP4 formats, plus a base model, post-trained variants, and a full training recipe in the NVIDIA-NeMo Nemotron GitHub repository. The release was positioned as a hardware strategy, timed two weeks before NVIDIA's fiscal Q2 results, with data center revenue reported at $89 billion, up 117% year-over-year per CNBC coverage.

Layer3Labs' benchmark guide reproduces NVIDIA's four headline evaluation scores for the Nemotron 3.5 Lightning BF16 checkpoint: SWE-bench Verified 51.56, GPQA Diamond 75.44, MMLU Pro 81.94, and PinchBench 85.37. These measurements position the 30B-total, 3B-active MoE model for coding and agent work rather than general chat. The explainer walks through what each score means in practice, noting that a low-50s SWE-bench Verified result is strong for the model class and signals the ability to read a codebase, locate bugs, and ship working patches. The guide also contextualizes NVIDIA's speed and efficiency claims against gpt-oss-120b and Qwen3.6-35B, citing the NVIDIA Developer Blog and the build.nvidia.com model card.

NVIDIA launched Nemotron 3.5 Lightning on August 11, 2026 as an open 30B-parameter Mixture-of-Experts model that activates only 3B parameters per token, built for high-volume tool-calling inside AI agent loops. The HokAI model page reports a 1,000,000-token context window, a PinchBench agentic score of 85.37, and day-zero inference support landing in vLLM, SGLang, and NeMo Switchyard on launch day. The page describes a hybrid LatentMoE architecture that interleaves Mamba-2 state-space layers, MoE experts, and select Attention layers, with tokens routing through a compressed latent dimension before expert computation. HokAI also lists a blended price of $0.088 per 1M tokens, 295 tok/s output speed, 51.6% on SWE-bench Verified, and 75.4% on GPQA Diamond from NVIDIA-published sources.