Nemotron 3.5 Lightning and the logic of the «middle layer» for agentic efficiency
On 11 August 2026, NVIDIA announced on its technical blog the release of Nemotron 3.5 Lightning, a mixture-of-experts model with 30 billion total parameters and 3 billion active per token. The official model card describes a hybrid architecture combining Mamba-2, MoE and attention, distributed under the OpenMDW 1.1 licence together with training data and recipes. According to the documentation, the model is intended primarily for customisation and post-training, with NVIDIA recommending the NVFP4 checkpoint for production deployment. On that note, the figures NVIDIA states for BF16 (MMLU Pro 81.94; SWE-bench Verified 51.56) do not match those reported by independent coverage for NVFP4 (81.62 and 52.80): how much the precision recommended for production affects the scores is still an open question.
On performance, the company claims generation speeds up to four times faster than comparably sized models and around 86% accuracy on the PinchBench agentic benchmark, completing 10,000 tasks 30% faster than Qwen3.6-35B at roughly the same accuracy. It is worth noting that these measurements come straight from the maker of the hardware the model runs on, and that the PinchBench comparison involves a single competing model picked by NVIDIA: no independent verification is available so far, and the blog specifies neither the hardware setup nor the parameters of the comparison, leaving real-world figures open to doubt.
One sticking point is memory. Although the architecture in theory supports a context length of up to 1 million tokens, availability drops to 256K tokens on a single H100 GPU because of physical limits. Single-card deployment requires 80 GB of video memory. Alongside the model, NVIDIA released NeMo Switchyard, a routing library designed to send each step of an agentic workflow to the most suitable model, putting into practice the strategy of leaving complex reasoning to larger models while routing the many repetitive steps to small, fast ones like Nemotron.
NVIDIA's move looks less like a head-on challenge to frontier models and more like an attempt to sell infrastructural efficiency. Offering a permissively licensed model — OpenMDW 1.1 allows commercial use with no fee, but it is a recent licence and far less road-tested than Apache 2.0, so read it before treating the two as equivalent — that runs well on existing hardware is a clever way to tie developers to the hardware the company sells. As long as the speed figures remain measurements taken by the party selling the GPUs, the claimed advantage is a commercial promise, not a verified result.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- Opened the official model card on Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16) and confirmed the hybrid Mamba-2 + MoE + attention architecture, 30B total and 3B active parameters, context up to 1M with a practical ceiling of 256K on a single H100, the OpenMDW 1.1 licence and the 11 August 2026 release date. NVIDIA's official technical blog confirms the NVFP4 and BF16 checkpoints, the distribution channels (Hugging Face, ModelScope, build.nvidia.com), the up-to-4x speed claim, the 86% PinchBench figure, the 10,000 tasks completed 30% faster than Qwen3.6-35B, and what NeMo Switchyard does. Independent corroboration from MarkTechPost (11 August), which reports the same numbers and attributes them explicitly to NVIDIA, plus the NIM API reference page confirming the model's availability. Set aside: the Apple–Alibaba story (anonymous sources, no public confirmation from either company), the MAI-Thinking-1 public preview (the model was already presented in June; the August announcement only concerns availability), and the FLI safety index and the DeepMind study on manipulation (published in July and in March–April 2026 respectively, re-dated by an aggregator).
- Incertezze
- Every speed and accuracy number is a measurement by the company that sells the GPUs the model runs on: there is no independent verification of the PinchBench results or of the 4x multiplier, and the blog does not detail the measurement setup (batch size, precision, GPU) for the comparison against «comparably sized» models. On PinchBench the comparison involves a single competing model, chosen by NVIDIA. The 1-million-token window is claimed at the architecture level, but on a single H100 it falls to 256K, so the advertised figure requires multiple GPUs. The OpenMDW 1.1 licence is recent and lightly tested compared with Apache 2.0: read it before assuming they are equivalent. It also remains unclear how much the official recommendation to use NVFP4 in production affects the scores, given that the two precisions do not produce the same numbers.
- Perché pubblicarla
- It is the most substantial open release of the week, and it moves the conversation to where it matters for people actually running agents: not peak intelligence, but the cost and latency of repetitive steps, with weights, data and recipes downloadable under a commercial licence and a single-GPU hardware profile. That the release comes from the GPU maker itself, with its own benchmarks and a router that orchestrates model selection, is precisely why the numbers deserve careful reading rather than repetition.