Nemotron 3.5 Lightning 30B A3B

Nemotron 3.5 Lightning 30B A3B

Quick answer: NVIDIA Nemotron 3.5 Lightning 30B A3B is a hybrid Mamba-2 + Mixture-of-Experts + Attention model with 30B total / 3B active parameters and a context window up to 1M tokens. Released August 11, 2026 under the OpenMDW-1.1 license, it's built for long-running autonomous agents and efficient single-GPU deployment (1× DGX Spark or 1× H100). NVIDIA's own harness measured 51.6% on SWE-bench Verified, 83.4 on PinchBench, and 75.6% on GPQA Diamond for the NVFP4 checkpoint covered on this page.

At a Glance

Where Nemotron 3.5 Lightning leads

• 1M-token context window with DSpark/DFlash speculative decoding and Multi-Token Prediction (MTP) for faster generation
• Runs on a single DGX Spark (GB10) or single H100 — no multi-GPU cluster required
• 83.4 on PinchBench (real-world OpenClaw agent tasks) — competitive with much larger frontier models on practical agent workflows
• Open weights, training data, and recipes under OpenMDW-1.1; commercial use permitted

Where it lags

• Only 3B active parameters caps raw reasoning ceiling vs. larger dense/MoE frontier models — HLE (text-only) sits at 11.7%
• τ³-bench (Banking) score of 9.3–9.5% shows financial-domain agentic tasks remain hard for this size class
• NVFP4-quantized checkpoint trades a small accuracy delta vs. the BF16 checkpoint on some benchmarks (e.g. SWE-bench Verified 51.6% NVFP4 vs. 52.8% BF16)

Best for: Long-running autonomous agents, sub-agent "workhorse" deployments inside larger agent systems, and local inference on personal hardware (DGX Spark, RTX 5090).

What Nemotron 3.5 Lightning 30B A3B Is

Nemotron 3.5 Lightning is NVIDIA's latest efficiency-focused release in the Nemotron family, built on a hybrid architecture that interleaves Mamba-2 state-space layers, Mixture-of-Experts layers, and a small number of Attention layers. Of its 30B total parameters, only about 3B are active per token — the "A3B" designation — which is what lets a model this size serve on a single H100 or a single DGX Spark (GB10) unit rather than requiring a multi-GPU cluster.

The release ships alongside three speculative-decoding strategies: DSpark (a semi-autoregressive drafter recommended for DGX Spark and low-concurrency data-center workflows), DFlash (a block-diffusion drafter), and MTP (Multi-Token Prediction, trained directly into the base model). NVIDIA reports DSpark delivering roughly a 15% output-throughput improvement over target-only decoding in independent community testing on DGX Spark hardware.

Pre-training used more than 20 trillion tokens with an NVFP4 recipe, and the model supports a context window of up to 1M tokens. Post-training combined supervised fine-tuning with multi-environment GRPO reinforcement learning across math, code, science, instruction-following, multi-step tool use, and structured-output environments. The NVFP4 checkpoint covered on this page is the result of a further post-training-quantization pass (Model Optimizer, four-over-six NVFP4 with FP8 KV cache).

NVIDIA positions the model explicitly for agentic use cases — long-running autonomous agents and "sub-agent workhorse" roles within larger multi-agent systems — rather than as a general chat-first flagship. Independent testing on DGX Spark (via vLLM 0.27.1) confirmed the deployment path works end-to-end, including nemotron_v3 reasoning-trace separation and native qwen3_coder-style tool calls.

Specifications

Pricing

Open weights under the OpenMDW-1.1 license — no vendor API pricing. Self-hosted inference cost depends on the deployment target (single DGX Spark, single H100, or multi-GPU H100/GB200 configurations); NVIDIA publishes reference vLLM, TensorRT-LLM, and SGLang serving recipes for each.

Public Benchmark Scores

Scores reported by NVIDIA under a consistent internal harness (NeMo Gym / NeMo Evaluator SDK); they may differ from other vendors' self-reported numbers on the same benchmarks. Not Benchgen measurements.

NVIDIA also reports SWE-bench Multilingual (39.3% NVFP4 / 36.5% BF16) — this benchmark does not yet have a Benchgen page, so it is noted here rather than in the scored leaderboard tables above.

How Nemotron 3.5 Lightning Performs on Agent Tasks

The most distinctive result in NVIDIA's own evaluation suite is PinchBench, a real-world OpenClaw coding-agent benchmark spanning productivity, research, writing, coding, analysis, email, and memory tasks rather than synthetic academic exercises. At 83.4–85.4, Nemotron 3.5 Lightning lands well above many larger models on PinchBench's public leaderboard — a meaningfully different signal than its more modest Humanity's Last Exam (10.5–11.7%) or τ³ Banking (9.3–9.5%) scores, which test narrower academic or single-domain capability rather than end-to-end agent task completion.

That gap is exactly the kind of thing Benchgen's approach is built to surface: a model can look unremarkable on knowledge benchmarks while still being a strong choice for a specific production agent harness. Community testing on DGX Spark independently confirmed the deployment path (vLLM 0.27.1, DSpark speculative decoding, nemotron_v3 reasoning parser, qwen3_coder tool calling) works end-to-end, with DSpark giving roughly a 15% throughput uplift over target-only decoding at low concurrency — relevant context for anyone evaluating this model as a sub-agent workhorse rather than judging it purely on static leaderboard scores.

Nemotron 3.5 Lightning vs Alternatives

Nemotron 3.5 Lightning is NVIDIA's newest 30B-class agent model, trading a slightly smaller active-parameter budget than Nemotron 3 Nano for a hybrid Mamba/MoE/Attention design, native speculative decoding, and a much stronger showing on real-world agent evaluation (PinchBench) than static academic benchmarks alone would suggest.

Related Models & Benchmarks

• Compare: Nemotron 3 Super 120B A12B · Nemotron 3 Nano 30B A3B · Nemotron 3 Ultra 550B A55B
• Benchmarks: PinchBench · SWE-bench Verified · MMLU-Pro · τ³ Banking

Frequently Asked Questions

Specs and benchmark scores from NVIDIA's Nemotron 3.5 Lightning 30B A3B NVFP4 model card (Hugging Face, released August 11, 2026) and the DGX Spark / GB10 developer forum thread. Last updated 2026-08-11.